Concerning the challenges of the diversity of data sources, non-independent and identical distribution of data and the heterogeneity of both computing capabilities and energy consumption of edge devices in Internet of Things (IoT), a computation offloading policy in Mobile Edge Computing (MEC) network that deploys both centralized learning and federated learning was proposed. Firstly, a system model of computation offloading related to both centralized learning and federated learning was built, considering the network transmission delay, computation delay and energy consumption of centralized learning and federated learning models. Then, with the system delay minimization as optimization object, considering the constraints of energy consumption and the training times based on machine learning accuracy, a computation offloading optimization model for machine learning was constructed. After that, the game for this computation offloading was formulated and analyzed. Based on the analysis results, an Energy-Constrained Delay-Greedy (ECDG) algorithm was proposed, which found the optimal solutions for the model via a two-stage policy of greedy decision and energy-constrained decision updating. Compared to the centralized-greedy and Federated Learning with Client Selection (FedCS) algorithms, ECDG algorithm has the lowest average learning delay, which is 1/10 of that in the centralized-greedy algorithm, and 1/5 of that in the FedCS algorithm. The experimental results show that, ECDG algorithms can automatically select the optimal machine learning models by computation offloading so that it can efficiently reduce the average machine learning delay, improve the energy efficiency of edge devices and satisfy the Quality of Service (QoS) requirements of IoT applications.

The real-time monitoring data generated by high-speed trains during running are usually processed by variable-length coding compression technology, which is convenient for transmission and storage. However, this method will complicate the internal structure of the compressed data, so that the corresponding data decompression process must follow the composition order of the compressed data, which is inefficient. In order to improve the decompression efficiency of high-speed train monitoring data, a parallel decompression algorithm for high-speed train monitoring data was proposed with the help of the speculation technology. Firstly, the structural characteristics of high-speed train monitoring data were studied, and the internal dependence that affects data division was analyzed. Secondly, the speculation technology was used to clean up internal dependence, and then, the data were divided into different parts tentatively. Thirdly, the division results were decompressed in a distributed computing environment in parallel. Finally, the parallel decompression results were combined together. Through this way, the decompression efficiency of high-speed train monitoring data was improved. Experimental results showed that on the computing cluster composed of 7 computing nodes, compared with the serial algorithm, the speedup of the proposed speculative parallel algorithm was about 3, showing a good performance of this algorithm. It can be seen that this algorithm can improve the monitoring data decompression efficiency significantly.

The task of image captioning is an important branch of image understanding. It requires not only the ability to correctly recognize the image content, but also the ability to generate grammatically and semantically correct sentences. The traditional encoder-decoder based model cannot make full use of image features and has only a single decoding method. In response to these problems, a multi-layer encoding and decoding model for image captioning based on attention mechanism named MLED was proposed. Firstly, Faster Region-based Convolutional Neural Network (Faster R-CNN) was used to extract image features. Then, Transformer was employed to extract three kinds of high-level features of the image. At the same time, the pyramid fusion method was used to effectively fuse the features. Finally, three Long Short-Term Memory (LSTM) Networks were constructed to decode the features of different layers hierarchically. In the decoding part, the soft attention mechanism was used to enable the model to pay attention to the important information required at the current step. The proposed model was tested on MSCOCO dataset and evaluated by BLEU, METEOR, ROUGE-L and CIDEr. Experimental results show that on the indicators BLEU-4, METEOR and CIDEr, the model is increased by 2.5 percentage points, 2.6 percentage points and 8.8 percentage points compared to the Recall what you see (Recall) model respectively, and is improved by 1.2 percentage points, 0.5 percentage points and 3.5 percentage points compared to the Hierarchical Attention-based Fusion (HAF) model respectively. The visualization of the generated description sentences show that the sentence generated by the proposed model can accurately reflect the image content.

Knowledge Graph (KG) strongly support the research of knowledge-driven artificial intelligence. Aiming at this fact, the existing technologies of knowledge graph and knowledge hypergraph were analyzed and summarized. At first, from the definition and development history of knowledge graph, the classification and architecture of knowledge graph were introduced. Second, the existing knowledge representation and storage methods were explained. Then, based on the construction process of knowledge graph, several knowledge graph construction techniques were analyzed. Specifically, aiming at the knowledge reasoning, an important part of knowledge graph, three typical knowledge reasoning approaches were analyzed, which are logic rule-based, embedding representation-based, and neural network-based. Furthermore, the research progress of knowledge hypergraph was introduced along with heterogeneous hypergraph. To effectively present and extract hyper-relational characteristics and realize the modeling of hyper-relation data as well as the fast knowledge reasoning, a three-layer architecture of knowledge hypergraph was proposed. Finally, the typical application scenarios of knowledge graph and knowledge hypergraph were summed up, and the future researches were prospected.

Fast and reliable hand pose estimation has a wide application in the fields such as human-computer interaction. In order to deal with the influences to the hand pose estimation caused by the light intensity changes, self-occlusions and large pose variations, a deep network framework based on label distribution learning was proposed. In the network, the point cloud of the hand was used as the input data, which was normalized through the farthest point sampling and Oriented Bounding Box (OBB). Then, the PointNet++ was utilized to extract features from the hand point cloud data. To deal with the highly non-linear relationship between the point cloud and the hand joint points, the positions of the hand joint points were predicted by the label distribution learning network. Compared with the traditional depth map based approaches, the proposed method was able to effectively extract discriminative hand geometric features with low computation cost and high accuracy. A set of tests were conducted on the public MSRA dataset to verify the effectiveness of the proposed hand pose estimation network. Experimental results showed that the average error of the hand joints estimated by this network was 8.43 mm, the average processing time of a frame was 12.8 ms, and the error of pose estimation was reduced by 11.82% and 0.83% respectively compared with the 3D CNN and Hand PointNet.

The histopathological image is the gold standard for identifying breast cancer, so that the automatic and accurate classification of breast cancer histopathological images is of great clinical application. In order to improve the classification accuracy of breast cancer histopathology images and thus meet the needs of clinical applications, a high-precision breast classification method that incorporates spatial and channel features was proposed. In the method, the histopathological images were processed by using color normalization and the dataset was expanded by using data enhancement, and the spatial feature information and channel feature information of the histopathological images were fused based on the Convolutional Neural Network (CNN) models DenseNet and Squeeze-and-Excitation Network (SENet). Three different BCSCNet (Breast Classification fusing Spatial and Channel features Network) models, BCSCNetⅠ, BCSCNetⅡ and BCSCNetⅢ, were designed according to the insertion position and the number of Squeeze-and-Excitation (SE) modules. The experiments were carried out on the breast cancer histopathology image dataset (BreaKHis), and through experimental comparison, it was firstly verified that color normalization and data enhancement of the images were able to improve the classification accuracy of breast canner, and then among the three designed breast canner classification models, the one with the highest precision was found to be BCSCNetⅢ. Experimental results showed that BCSCNetⅢ had the accuracy of binary classification ranged from 99.05% to 99.89%, which was improved by 0.42 percentage points compared with Breast cancer Histopathology image Classification Network (BHCNet); and the accuracy of multi-class classification ranged from 93.06% to 95.72%, which was improved by 2.41 percentage points compared with BHCNet. It proves that BCSCNet can accurately classify breast cancer histopathological images and provide reliable theoretical support for computer-aided breast cancer diagnosis.

The various shapes of ferrite magnetic tiles and the wide varieties of their surface defects are great challenges for computer vision based surface defect quality recognition. To address this problem, the deep learning technique was introduced to the magnetic tile surface quality recognition, and a surface defect detection system for magnetic tiles was proposed based on convolution neural networks. Firstly, the tile target was segmented from the collected image and was rotated in order to obtain the standard image. After that, the improved multiscale ResNet18 was used as the backbone network to design the recognition system. During the training process, a novel within-class mixup operation was designed to improve the generalization ability of the system on the samples. To close to the practical application scenes, a surface defect dataset was built with the consideration of illumination changes and posture differences. Experimental results on the self-built dataset indicate that the proposed system achieves recognition accuracy of 97.9%, and provides a feasible idea for the automatic recognition of magnetic tile surface defects.

Concerning the problems of 3D point cloud data processing and on-road obstacle detection based on Light Detection And Ranging (LiDAR), a deep learning-based on-road obstacle detection method was proposed. First, the statistical filtering algorithm was applied to eliminate the outliers from the original point cloud, improving the roughness of point clouds. Then, an end-to-end deep neural network named VNMax was proposed, the max pooling was used to optimize the structure of Region Proposal Network (RPN), and an improved target detection layer was built. Finally, training and testing experiments were performed on KITTI dataset. The results show that, by filtering, the average distance between the points in point cloud is reduced effectively. For the car location processing results of easy, medium difficult and hard detection tasks in KITTI dataset, it can be seen that the average precisions of the proposed method are improved by 11.30 percentage points, 6.02 percentage points and 3.89 percentage points, respectively, compared with those of the VoxelNet. Experimental results show that the statistical filtering algorithm is still an effective 3D point cloud data processing method, and the max pooling module can improve the learning performance and object location ability of the deep neural network.

In the ship pipeline layout at three-dimensional environment, aiming at the problems that there are too many constraints, the engineering rules are difficult to quantify and the appropriate optimization evaluation function is hard to determine, a new ship pipeline automatic layout method was proposed. Firstly, the hull and ship equipments were simplified by the Aixe Align Bounding Box (AABB) method, which means that they were discretized into space nodes, and the initial pheromones and energy values of them were given, the obstacles in the space were marked, and the specific quantitative forms for the main pipe-laying rules were given. Secondly, with the combination of Rapidly-exploring Random Tree (RRT) algorithm and Ant Colony Optimization (ACO) algorithm, the direction selection strategy, obstacle avoidance strategy and variable step strategy were introduced to improve the search efficiency and success rate of the algorithm, and then the ACO algorithm was used to optimize the path iteratively by establishing the optimization evaluation function, so as to obtain the comprehensive optimal solution that meets the engineering rules. Finally, the computer simulated cabin space layout environment was used to carry out automatic pipe-laying simulation experiments, which verified the effectiveness and practicability of the proposed method.

Optical flow method is an important and effective deformation registration algorithm based on optical flow field model. Aiming at the problem that the feature quality used by the existing optical flow method is not high enough to make the registration result accurate, combining the features of deep convolutional neural network and optical flow method, a deformable medical image registration algorithm based on Deep Convolution Feature Based Optical Flow (DCFOF) was proposed. Firstly, the deep convolution feature of the image block where each pixel in the image was located was densely extracted by using a deep convolutional neural network, and then the optical flow field was solved based on the deep convolution feature difference between the fixed image and the floating image. By extracting more accurate and robust deep learning features of the image, the optical flow field obtained was closer to the real deformation field, and the registration accuracy was improved. Experimental results show that the proposed algorithm can solve the problem of deformable medical image registration effectively, and has the registration accuracy better than those of Demons algorithm, Scale-Invariant Feature Transform（SIFT） Flow algorithm and professional registration software of medical images called Elastix.

In order to solve the interference of similar background to object tracker in ground battlefield environment, an object tracking algorithm combining correlation filtering and improved color probability model was proposed. Firstly, based on the traditional color probability model, a color probability model emphasizing foreground was proposed by using the difference between foreground object histogram and background histogram. Then, a spatial penalty matrix was generated according to the correlation filter response confidence and maximum response position. This matrix was used to punish the likelihood probability of background pixel determined by the correlation filter, and the response map of the color probability model was obtained by using the method of integral image. Finally, the response maps obtained by the correlation filter and the color probability model were fused, and the maximum response position of the fusion response map was the central position of the object. The proposed algorithm was compared with 5 state-of-the-art algorithms such as Circulant Structure of tracking-by-detection filters with Kernels (CSK), Kernelized Correlation Filters (KCF), Discriminative Scale Space Tracking (DSST), Scale Adaptive Multiple Feature (SAMF) and Staple in tracking performance. The experimental results on OTB-100 standard dataset show that, the proposed algorithm has the overall accuracy improved by 3.06% to 55.98%, and the success rate improved by 2.24% to 54.97%； and under similar background interference, the proposed algorithm has the accuracy improved by 10.28% to 43.9%, and the success rate improved by 8.3% to 48.29%. The experimental results on 36 battlefield video sequences show that, the proposed algorithm has the overall accuracy improved by 2.2% to 45.98%, and the success rate improved by 3.01% to 58.27%. It can be seen that the proposed algorithm can better deal with the interference of similar background in the ground battlefield environment, and provide more accurate position information for the weapon platform.

An Ultra WideBand (UWB)/ Pedestrian Dead Reckoning (PDR) fusion positioning algorithm with adaptive coefficient adjustment based on UWB error prediction was proposed in order to improve the UWB performance and reduce the PDR accumulative errors in the indoor Non-Line-Of-Sight (NLOS) positioning scenes and solve the UWB performance degradation caused by environmental factors. On the basis of the creative proposal of predicting the UWB positioning errors in complex environment by Support Vector Machine (SVM) regression model, UWB/PDR fusion positioning performance was improved by adding adaptive adjusted parameters to the conventional Extended Kalman Filter (EKF) algorithm. The experimental results show that the proposed algorithm can effectively predict the current UWB positioning errors in the complex UWB environment, and increase the accuracy by adaptively adjusting the fusion parameters, which makes the positioning error reduced by 18.2% in general areas and reduced by 48.7% in the areas with poor UWB accuracy compared with those of the conventional EKF algorithm, so as to decrease the environmental impact on the UWB performance. In complex scenes of both Line-Of-Sight (LOS) and NLOS including UWB, the positioning error per 100 meters is reduced from meter scale to decimeter scale, which reduces the PDR errors in NLOS scenes.

In order to avoid the flower pollination algorithm falling into local optimum in the identification process of the extreme learning machine, an extreme learning machine algorithm based on cloud quantum flower pollination was proposed. Firstly, cloud model and quantum system were introduced into the flower pollination algorithm to enhance the global search ability of the flower pollination algorithm, so that the particles were able to perform optimization in different states. Then, the cloud quantum flower pollination algorithm was used to optimize the parameters of the extreme learning machine in order to improve the identification accuracy and efficiency of the extreme learning machine. In the experiments, six benchmark functions were used to simulate and compare several algorithms. It is verified by the comparison results that the performance of proposed cloud quantum flower pollination algorithm is superior to those of other three swarm intelligence optimization algorithms. Finally, the improved extreme learning machine algorithm was applied to the identification of oil and gas layers. The experimental results show that, the identification accuracy of the proposed algorithm reaches 98.62%, and compared with the classic extreme learning machine, its training time is shortened by 1.680 2 s. The proposed algorithm has high identification accuracy and efficiency, and can be widely applied to the actual classification field.

For the parallel machine scheduling problem of minimizing the maximum completion time, an Improved Discrete Artificial Bee Colony algorithm (IDABC) was proposed by considering the processing efficiency of the machine and the delivery time of the product as well as introducing the mathematical model of the problem. Firstly, a uniformly distributed population and a generation strategy of the parameters to be optimized were achieved by adopting the population initialization strategy, resulting in the improvement of the convergence speed of population. Secondly, the mutation operator in the differential evolution algorithm and the idea of simulated annealing algorithm were used to improve the local search strategy for the employed bee and the following bee, and the scout bee was improved by using the high-quality information of the optimal solution, resulting in the increasement of the population diversity and the avoidance of trapping into the local optimum. Finally, the proposed algorithm was applied in the parallel machine scheduling problem to analyze the performance and parameters of the algorithm. The experimental results on 15 examples show that compared with the Hybrid Discrete Artificial Bee Colony algorithm (HDABC), IDABC has the accuracy and stability improved by 4.1% and 26.9% respectively, and has better convergence, which indicates that IDABC can effectively solve the parallel machine scheduling problem in the actual scene.

Only one semantic vector can be generated by word-embedding technologies such as Word2vec or GloVe for polysemous word. In order to solve the problem, a sentiment analysis model based on ELMo (Embedding from Language Model) and Multi-Scale Convolutional Neural Network (MSCNN) was proposed. Firstly, ELMo model was used to learn the pre-training corpus and generate the context-related word vectors. Compared with the traditional word embedding technology, in ELMo model, word features and context features were combined by bidirectional LSTM (Long Short-Term Memory) network to accurately express different semantics of polysemous word. Besides, due to the number of Chinese characters is much more than English characters, ELMo model is difficult to train for Chinese corpus. So the pre-trained Chinese characters were used to initialize the embedding layer of ELMo model. Compared with random initialization, the model training was able to be faster and more accurate by this method. Then, the multi-scale convolutional neural network was applied to secondly extract and fuse the features of word vectors, and generate the semantic representation for the whole sentence. Experiments were carried out on the hotel review dataset and NLPCC2014 task2 dataset. The results show that compared with the attention based bidirectional LSTM model, the proposed model obtain 1.08 percentage points improvement of the accuracy on hotel review dataset, and on NLPCC2014 task2 dataset, the proposed model gain 2.16 percentage points improvement of the accuracy compared with the hybrid model based on LSTM and CNN.

To address the problem that the results of text-to-image synthesis tasks have wrong target structures and unclear image textures, a Multi-level Progressive Resolution Generative Adversarial Network (MPRGAN) model was proposed based on Attentional Generative Adversarial Network (AttnGAN). Firstly, a semantic separation-fusion generation module was used in low-resolution layer, and the text feature was separated into three feature vectors by the guidance of self-attention mechanism and the feature vectors were used to generate feature maps respectively. Then, the feature maps were fused into low-resolution map, and the mask images were used as semantic constraints to improve the stability of the low-resolution generator. Finally, the progressive resolution residual structure was adopted in high-resolution layers. At the same time, the word attention mechanism and pixel shuffle were combined to further improve the quality of the generated images. Experimental results showed that, the Inception Score (IS) of the proposed model reaches 4.70 and 3.53 respectively on datasets of Caltech-UCSD Birds-200-2011 (CUB-200-2011) and 102 category flower dataset (Oxford-102), which are 7.80% and 3.82% higher than those of AttnGAN, respectively. The MPRGAN model can solve the instability problem of structure generation to a certain extent, and the images generated by the proposed model is closer to the real images.

The current consistency detection methods of software design and implementation require a large number of template sets and are difficult to generalize. In order to solve these problems, a consistency analysis method of software design and implementation based on control flow was proposed. Firstly, the pseudocode of the design document and the source code of the program were converted into the intermediate representations with the same features, and the design feature and the implementation feature were respectively extracted from the intermediate representations. The features include the function call relationship which can reflect the system structure and the control flow information which can reflect the internal structure of the function. Then, the design feature model and the implementation feature model were respectively established according to the design feature and the implementation feature. Finally, the similarity of the feature model was measured by calculating the feature similarity, so as to obtain the consistency detection result. Experimental results show that this method can correctly detect the inconsistent function call relationship when the function call relationship realized by the software is inconsistent with the design, and can correctly detect the inconsistency of the internal structure of the function when the function call relationship realized by the software is consistent with the design, with the accuracy reached 92.85%. This method can effectively obtain the consistency detection results without any template set, and has superior generality.

The traditional software evaluation methods are lack of consideration for user requirements. For this problem, an evaluation model of software quality based on user group decision-making was proposed. Firstly, it is found that the existing projection measure is not always reasonable in real number and interval vector spaces. Therefore, a new normalized projection measure was proposed to comprehensively measure the proximity between two vectors or matrices and the measure allows the evaluation matrix with hybrid decision-making information. Secondly, the new projection measure was fused in the improved Technique for Order Performance by Similarity to Ideal Solution (TOPSIS) technique. On this basis, a new group decision-making model with hybrid information of real number and interval was developed. And the pseudocode of algorithm was provided. Finally, the new model was applied to software quality evaluation. The requirements of software users were focused and the evaluation information of users group was synthesized in this method. The effectiveness and feasibility of the proposed method were illustrated by a practical example of software quality comprehensive evaluation and the experimental analysis.

Aiming at the inefficiency of neural network inferential calculation executed by mobile intelligent devices using ARM processor, a set of Single precision floating GEneral Matrix Multiply (SGEMM) algorithm optimization scheme based on ARMv8 architecture was proposed. Firstly, it was determined that the computational efficiency of the processor based on ARMv8 architecture executing SGEMM algorithm was limited by the vectorized computation unit usage scheme, the instruction pipeline, and the probability of occurrence of cache miss. Secondly, three optimization techniques:vector instruction inline assembly, data rearrangement and data prefetching were implemented for the three reasons that the computational efficiency was limited. Finally, the test experiments were designed based on three matrix patterns commonly used in the neural network of speech direction and the programs were run on the RK3399 hardware platform. The experimental results show that, the single-core computing speed is 10.23 GFLOPS in square matrix mode, reaching 78.2% of the measured floating-point peak value; the single-core computing speed is 6.35 GFLOPS in slender matrix mode, reaching 48.1% of the measured floating-point peak value; and the single-core computing speed is 2.53 GFLOPS in continuous small matrix mode, reaching 19.2% of the measured floating-point peak value. With the optimized SGEMM algorithm deployed into the speech recognition neural network program, the actual speech recognition speed of program is significantly improved.

To solve the signal demodulation problem of asymmetric Paired Carry Multiple Access (PCMA) composed of the same frequency of main station and small station signals, a framework to realize this kind of signal demodulation was constructed. Parameter estimation is an indispensable part in the realization of two-way signal separation and demodulation for asymmetric PCMA communication systems. For the estimation accuracy of amplitude parameters, a searching amplitude estimation algorithm based on fourth-power method was proposed. Firstly, the demodulation model for asymmetric PCMA systems was established and the basic assumptions were made. Then the phase errors under different assumptions were compared with each other and the influence of phase error on the amplitude estimation algorithm was analyzed. Finally, a new amplitude estimation algorithm was proposed. Experimental results show that, under same Signal-to-Noise Ratio (SNR), the demodulation performance of the small station signal under normal phase error is inferior to its demodulation performance under mean value condition. When the order of magnitude of the Bit Error Rate (BER) is 10 ^-4, the demodulation performance of small station signal is improved by 1 dB with the improved algorithm, proving that the improved algorithm is better than fourth-power method.

Aiming at the problem that traditional single detection algorithm models have low detection rate and slow detection speed on different types of attacks in industrial control system, an intrusion detection model combining optimized Support Vector Machine (SVM) and K-means++algorithm was proposed. Firstly, the original dataset was preprocessed by Principal Component Analysis (PCA) to eliminate its correlation. Secondly, an adaptive mutation process was added to Particle Swarm Optimization (PSO) algorithm to avoid falling into local optimal solution during the training process. Thirdly, the PSO with Adaptive Mutation (AMPSO) algorithm was used to optimize the kernel function and penalty parameters of the SVM. Finally, a K-means algorithm improved by density center method was united with the optimized support vector machine to form the intrusion detection model, achieving anomaly detection of industrial control system. The experimental results show that the proposed method can significantly improve the detection speed and the detection rate of various attacks.

To solve the problems that group effect is neglected when selecting dictionary bases in sparse coding models, and distance between a features and a dictionary base can not be effectively measured by Euclidean distance, Non-negative Local Sparse Coding algorithm based on Elastic net and Histogram intersection (EH-NLSC) was proposed. Firstly, with elastic-net model introduced in the optimization function to remove the restriction on selected number of dictionary bases, multiple groups of correlation features were selected and redundant features were eliminated, improving the discriminability and effectiveness of the coding. Then, histogram intersection was introduced in the locality constraint of the coding, and the distance between the feature and the dictionary base was redefined to ensure that similar features share their local bases. Finally, multi-class linear Support Vector Machine (SVM) was adopted to realize image classification. The experimental results on four public datasets show that compared with LLC (Locality-constrained Linear Coding for image classification) and NENSC (Non-negative Elastic Net Sparse Coding), the classification accuracy of EH-NLSC is increased by 10 percentage points and 9 percentage points respectively on average, proving its effectiveness in image representation and classification.

Artificial identification of ischemic stroke lesion is time-consuming, laborious and easy be added subjective differences. To solve this problem, an automatic segmentation algorithm based on 3D deep residual network and cascade U-Net was proposed. Firstly, in order to efficiently utilize 3D contextual information of the image and the solve class imbalance issue, the patches were extracted from the stroke Magnetic Resonance Image (MRI) and put into network. Then, a segmentation model based on 3D deep residual network and cascade U-Net was used to extract features of the image patches, and the coarse segmentation result was obtained. Finally, the fine segmentation process was used to optimize the coarse segmentation result. The experiment results show that, on the dataset of Ischemic Stroke LEsion Segmentation (ISLES), for the proposed algorithm, the Dice similarity coefficient reached 0.81, the recall reached 0.81 and the precision reached 0.81, the distance coefficient Average Symmetric Surface Distance (ASSD) reached 1.32 and Hausdorff Distance (HD) reached 22.67. Compared with 3D U-Net algorithm, level set algorithm, Fuzzy C-Means (FCM) algorithm and Convolutional Neural Network (CNN) algorithm, the proposed algorithm has better segmentation performance.

Aiming at the problem of poor pedestrian detection effect in complex scenes, a pedestrian detection algorithm based on improved Mask R-CNN framework was proposed with the use of the leading research results in deep learning-based object detection. Firstly, K-means algorithm was used to cluster the object frames of the pedestrian datasets to obtain the appropriate aspect ratio. By adding the set of aspect ratio (2:5), 12 anchors were able to be adapted to the size of the pedestrian in the image. Secondly, combined with the technology of fine-grained image recognition, the high accuracy of pedestrian positioning was realized. Thirdly, the foreground object was segmented by the Full Convolutional Network (FCN), and pixel prediction was performed to obtain the local mask (upper body, lower body) of the pedestrian, so as to achieve the fine-grained detection of pedestrians. Finally, the overall mask of the pedestrian was obtained by learning the local features of the pedestrian. In order to verify the effectiveness of the improved algorithm, the proposed algorithm was compared with the current representative object detection methods (such as Faster Region-based Convolutional Neural Network (Faster R-CNN), YOLOv2 and R-FCN (Region-based Fully Convolutional Network)) on the same dataset. The experimental results show that the improved algorithm increases the speed and accuracy of pedestrian detection and reduces the false positive rate.

To solve the problems of high complexity of precoding and difficulty of linear matrix inversion in downlink Massive Multi-Input Multi-Output (Massive MIMO) system, a precoding algorithm based on low-complexity Symmetric Successive Over Relaxation Preconditioned Conjugate Gradient (SSOR-PCG) was proposed. Based on preconditioned Conjugate Gradient Precoding (PCG) algorithm, a Symmetric Successive Over Relaxation (SSOR) algorithm was used to preprocess the matrix to reduce its condition number, accelerating the convergence speed and the decreasing the complexity. Simulation results demonstrate that compared with PCG algorithm, the proposed algorithm has running time of around 88.93% shortened and achieves convergence when the Signal-to-Noise Ratio (SNR) is 26 dB. Furthermore, compared to zero-forcing precoding algorithm, the proposed algorithm requires only two iterations capacity-approaching performance,the overall complexity is reduced by one order of magnitude, and the bit error rate is decreased by about 49.94%.

In the mobile robot path planning problem in static multi-obstacle environment, Particle Swarm Optimization (PSO) algorithm has the disadvantages of easy premature convergence and poor local optimization ability, resulting in low accuracy of robot path planning. To solve the problem, a Multi-Objective Grasshopper Optimization Algorithm (MOGOA) was proposed. The path length, smoothness and security were taken as path optimization targets according to the mobile robot path planning requirements, and the corresponding mathematical model of multi-objective optimization problem was established. In the process of population search, the curve adaptive strategy was introduced to speed up the convergence of the algorithm, and the Pareto optimal criterion was used to solve the coexistence problem of the above three targets. Experimental results show that the proposed algorithm finds shorter paths and shows better convergence while solving the above problems. Compared with the Multi-Objective Particle Swarm Optimization (MOPSO) algorithm, the proposed algorithm has the path length reduced by about 2.01 percentage, and the number of iterations reduced by about 19.34 percentage.

To solve traffic congestion problems, a new road condition judgment model was proposed. Firstly, the model was based on YOLOv3 target detection algorithm. Then, according to the eigenvalue matrix corresponding to the picture, the difference between adjacent frames was made by the eigenvalue matrix, and the difference value was compared with preset value to determine whether the current road was in a congested state or a normal traffic state. Secondly, the current calculated road state was compared with previous two calculated road states. Finally, the state statistics method in the model was used to calculate the duration of a state (congestion or patency) of road. The proposed model could analyze the states of three lanes of a road at the same time. Through experiments, the average accuracy of model to judge the state of single lane could reach 80% or more, and it was applicable to both day and night roads.

To solve the problems of external disturbances and the uncertainty of system model parameters for the underactuated quadrotor Unmanned Aerial Vehicle (UAV) existing in actual flight, a flight control scheme based on Extended State Observer (ESO) and integral backstepping sliding mode was designed. Firstly, according to the semi-coupling characteristics and the strict feedback architecture of system, a backstepping control was adopted to design the attitude inner loop and the position outer loop controllers. Then, a sliding mode algorithm with strong anti-jamming ability and integral control were incorporated to enhance system robustness and reduce static error respectively. Finally, ESO was used to eliminate the total internal and external disturbances and to compensate the interference in the control law online. The closed-loop control system was proven to be globally asymptotically stable by the Lyapunov stability analysis. In addition, the effectiveness and robustness of the proposed flight control scheme were verified through simulation analysis.

Focused on that limitation of single-layer structure on image feature learning ability, a deep architecture based on sparse representation of image blocks was proposed, namely Multi-layer incorporating Locality and non-negativity Laplacian Sparse Coding method (MLLSC). Each image was divided uniformly into blocks and SIFT (Scale-Invariant Feature Transform) feature extraction on each image block was performed. In the sparse coding stage, locality and non-negativity were added in the Laplacian sparse coding optimization function, dictionary learning and sparse coding were conducted at the first and second levels, respectively. To remove redundant features, Principal Component Analysis (PCA) dimensionality reduction was performed before the second layer of sparse coding. And finally, multi-class linear SVM (Support Vector Machine) was adopted for image classification. The experimental results on four standard datasets show that MLLSC has efficient feature expression ability, and it can capture deeper feature information of images. Compared with the single-layer algorithms, the accuracy of the proposed algorithm is improved by 3% to 13%; compared with the multi-layer sparse coding algorithms, the accuracy of the proposed algorithm is improved by 1% to 2.3%. The effects of different parameters were illustrated, which fully demonstrate the effectiveness of the proposed algorithm in image classification.

Since the vehicles provided by current online car rental platforms are in the fixed price, there are some issues coming up such as unreasonable allocation of the vehicle resources, unreliable price that could not indicates the real market supply and demand timely, and generally low social welfare. Therefore, an auction based vehicle allocation and pricing mechanism for car rental was proposed. Firstly, a mathematical model and a social welfare maximization objective function were established by studying the model of online car rental issues. Secondly, based on the minimum cost and maximum flow algorithm, the optimal vehicle resource allocation algorithm was adopted among the rental vehicle allocation algorithms. Finally, in terms of the price calculation algorithms, a truthful VCG (Vickrey-Clarke-Groves) price algorithm was used to calculate the final price. As a result, compared with the traditional first-come-first-serving algorithms, the order success rate of the proposed scheme was increased by 20% to 30%, and the revenue was increased by about 30%. Theoretical analysis and experiment results show that the proposed mechanism has the advantages of optimizing vehicle allocation and flexible price strategy.