Pedestrian re-identification tasks lack the consideration of the pedestrian feature scale variation during feature extraction, so that they are easily affected by environment and have low accuracy of pedestrian re-identification. In order to solve the problem, a pedestrian re-identification method based on multi-scale feature fusion was proposed. Firstly, in the shallow layer of the network, multi-scale pedestrian features were extracted through mixed pooling operation, which was helpful to improve the feature extraction capability of the network. Then, strip pooling operation was added to the residual block to extract the remote context information in horizontal and vertical directions respectively, which avoided the interference of irrelevant regions. Finally, after the residual network, the dilated convolutions with different scales were used to further preserve the multi-scale features, so as to help the model to analyze the scene structure flexibly and effectively. Experimental results show that, on Market-1501 dataset, the proposed method has the Rank1 of 95.9%, and the mean Average Precision (mAP) of 88.5%; on DukeMTMC-reID dataset, the proposed method has the Rank1 of 90.1%, and the mAP of 80.3%. It can be seen that the proposed method can retain the pedestrian feature information better, thereby improving the accuracy of pedestrian re-identification tasks.

In the advanced assisted driving device, machine vision technology was used to process the video of vehicles in front in real time to dynamically recognize and predict the posture and behavior of vehicle. Concerning low precision and large delay of this kind of recognition algorithm, a deep learning algorithm for vehicle behavior dynamic recognition based on Long Short-Term Memory (LSTM) was proposed. Firstly, the key frames in vehicle behavior video were extracted. Secondly, a dual convolutional network was introduced to analyze the feature information of key frames in parallel, and then LSTM network was used to sequence the extracted characteristic information. Finally, the output predicted score was used to determine the behavior type of vehicle. The experimental results show that the proposed algorithm has an accuracy of 95.6%, and the recognition time of a single video is only 1.72 s. The improved dual convolutional network algorithm improves the accuracy by 8.02% compared with ordinary convolutional network and increases by 6.36% compared with traditional vehicle behavior recognition algorithm based on a self-built dataset.

The self-driving mine locomotive needs to detect and locate pedestrians in front of it in the underground roadway in real-time. Non-visual methods such as laser radar are costly, while traditional visual methods based on feature extraction cannot solve the problem of poor illumination and uneven light in the laneway. To solve the problem, a pedestrian visual positioning algorithm for underground roadway based on deep learning was proposed. Firstly, the overall structure of the system based on deep learning network was given. Secondly, a multi-layer Convolutional Neural Network (CNN) for object detection was built to calculate the two-dimensional coordinates and the size of bounding box of pedestrians in visual field of the self-driving locomotive. Thirdly, the third-dimensional distance between the pedestrian in the image and the locomotive was calculated by polynomial fitting. Finally, the model was trained, verified and tested through real sample sets. Experimental results show that the accuracy of the proposed algorithm reaches 94%, the speed achieves 25 frames per second, and the distance detection error is less than 4%, thus efficient and real-time laneway pedestrian visual positioning is realized.

In order to solving the problem that the pedestrian detection algorithm based on Aggregated Channel Feature (ACF) has a low detection precision and a high false detection rate in complex scenes, a multi-channel pedestrian detection algorithm combined with features of texture and contour was proposed in this paper. The algorithm flows included training classifier and detection. In the training phase, the ACF, the texture features of Local Binary Patterns (LBP) and the contour features of Sketch Tokens (ST) were extracted, and trained separately by the Real AdaBoost classifier. In the detection phase, the cascading detection idea was used. The ACF classifier was used to deal with all objects, then the complicated classifier of LBP and ST were used to gradually filter the result of the previous step. In the experiment, the INRIA data set was used in the simulation of our algorithm, the results show that our algorithm achieves a Log-Average Miss Rate (LAMR) of 13.32%. Compared with ACF algorithm, LAMR is decreased by 3.73 percent points. The experimental results verify that LBP and ST can be used as a complementation of ACF. So some objects of false detection can be eliminated in the complicated scenes and the accuracy can be improved. At the same time, the efficiency of multi-feature algorithm is ensured by cascading detection.

Median Prior (MP) reconstruction algorithm combined with nonlocal means fuzzy diffusion and extended neighborhood bilateral filter was proposed to reduce the streak artifacts in low-dose Computed Tomography (CT) reconstruction. In the new algorithm, the nonlocal means fuzzy diffusion method was used to improve the median of the prior distribution Maximum A Posterior (MAP) reconstruction algorithm at first, which reduced the noise in the reconstruction image; then, the bilateral filtering method based on the expended neighborhood was applied to preserve the edges and details of the reconstruction image and improve the Signal-to-Noise Ratio (SNR). The Shepp-Logan model and the thorax phantom were used to test the effectiveness of the proposed algorithm. The experimental results show that the proposed method has the smaller values of the Normalized Mean Square Distance (NMSD) and Mean Absolute Error (MAE) and the highest SNR (10.20 dB and 15.51 dB, respectively) in the two experiment images, compared with Filtered Back Projection (FBP), Median Root Prior (MRP), NonLocal Mean MP (NLMMP) and NonLocal Mean Bilateral Filter MP (NLMBFMP) algorithms. The experimental results show that the proposed reconstruction algorithm can reduce noise while keeping the edges and details of the image, which improves the deterioration problem of the low-dose CT image and obtains the image with higher SNR and quality.

Aiming at false segmentation of small regions and high computational complexity in traditional color image segmentation algorithm, a hierarchical method of color image segmentation based on rough set and HIS (Hue-Saturation-Intensity) space was proposed. Firstly, for the reason that the singularities in HSI space are the achromatic pixels in RGB space, the achromatic regions of RGB space were segmented and labeled in order to remove the singularities from the original image. Secondly, the original image was converted from RGB space to HSI space. In intensity component, in view of spatial neighbor information and regional distribution difference, the original histogram was weighted by homogeneity function with changing thresholds and gradience. The weighted and original histograms were respectively used as the upper and lower approximation sets of rough set. The new roughness function was defined and applied to image segmentation. Then the different regions obtained in the previous stage were segmented according to the histogram in hue component. Finally, the homogeneous regions were merged in RGB space in order to avoid over-segmentation. Compared with the method based on rough set proposed by Mushrif etc. (MUSHRIF M M, RAY A K. Color image segmentation: rough-set theoretic approach. Pattern Recognition Letters, 2008, 29(4): 483-493), the proposed method can segment small regions easily, avoid the false segmentation caused by the correlation between RGB color components, and the executing speed is 5-8 times faster. The experimental results show the proposed method yields better segmentation, and it is efficient and robust to noise.

Most existing works on memory management focus on the efficiency, which are real-time, but have memory fragmentation problems. To address the problem, an efficient memory management algorithm based on segment tree was proposed. The proposed method built a memory management segment tree by dividing memory space into segments, and allocated and reclaimed memory efficiently and flexibly based on the memory management segment tree to reduce the memory fragmentation. Furthermore, a method was proposed to optimize the space complexity of segment trees. The experimental results show that the proposed method has advantages in terms of efficiency, memory fragmentation, storage space, and so on.

In order to make up for the limitation of research scenario of call admission control in heterogeneous wireless networks and reduce blindness of access network selection, the scenario was extended from integrated system with two networks to integrated system with multiple networks, and a joint call admission control algorithm based on reputation model was proposed. The reputation model was applied in the network selection and feedback mechanism. On the user side, the terminals chose the access network according to the networks' reputation; on the network side, the networks made decisions by adaptive bandwidth management and buffer queuing policy to enhance the probability of successful acceptation. Simulation results show that by using the proposed algorithm, new call blocking probability and handoff call dropping probability can be reduced effectively.

In a Virtual Machine (VM) computing environment, it is difficult to monitor and allocate the VM's memory in real-time. To overcome these shortcomings, a real-time method of monitoring and adjusting memory for Xen virtual machine called Xen Memory Monitor and Control (XMMC) was proposed and implemented. This method used hypercall of Xen, which could not only real-time monitor the VM's memory usage, but also dynamically real-time allocated the VM's memory by demand. The experimental results show that XMMC only causes a very small performance loss, less than 5%, to VM's applications. It can real-time monitor and adjust on demand VM's memory resource occupations, which provides convenience for the management of multiple virtual machines.

Because of its unique advantage of information propagation, the online social network has been a popular social communication platform. In view of the characteristics of the form of information propagation and the dynamics theory of infectious diseases, this paper put forward the model of information propagation through online social network. The model considered the influence of different users' behaviors on the transmission mechanism, set up the evolution equations of different user nodes, simulated the process of information propagation, and analyzed the behavior characteristics of the different types of users and main factors that influenced the information propagation. The experimental results show that different types of users have special behavior rules in the process of information propagation, i.e., information cannot be transported endlessly, and be reached at a stationary state, and the larger the spread coefficient or immune coefficient is, the faster it reached the stationary state.

Most of existing manifold learning algorithms are not capable of dealing with new arrival samples. Although some incremental algorithms are developed via extending a specified manifold learning algorithm, most of them have some disadvantages more or less. In this paper, a novel and more Generalized Incremental Manifold Learning algorithm based on local smoothness is proposed (GIML). GIML algorithm first extracts the local smoothness structure of data set via local PCA. Then the optimal linear transformation, which transforms the local smoothness structure of new arrival sample’s neighborhood to its corresponded low-dimensional embedding coordinates, is computed. Finally the low-dimensinal embedding coordinates of new arrival samples are obtained by the optimal transformation. Extensive and systematic experiments are conducted on both artificial and real image data sets. Experimental results demonstrate that our GIML algotithm is an effective incremental manifold learning algorithm and outperforms other existing algirthms.