Visual perception， as one of the core technologies of environmental understanding， provides accurate environmental information for intelligent mobile systems （such as autonomous driving） and is an important prerequisite for ensuring safety decisions. 3D object detection technology based on Bird’s Eye View （BEV） has become the mainstream paradigm in the field of environmental perception because of its efficiency and accuracy. To further promote the research of 3D object detection algorithms based on BEV， the following was performed. Firstly， the BEV 3D object detection algorithms were classified systematically， and according to the modals of the input data， they were divided into three categories： pure camera algorithms， pure LiDAR algorithms and camera-LiDAR fusion algorithms. Secondly， the role of pre-training algorithms in improving detection performance was explored. Thirdly， the advantages and disadvantages of the algorithms fusing temporal features in dynamic scenarios and the performance of the algorithms fusing height features in complex environments were analyzed. Fourthly， the breakthrough progress made by large model collaborative BEV object detection in object detection accuracy and scenario understanding was sorted out. Finally， the core conclusions of BEV 3D object detection algorithms were summarized， and future research directions were looked forward， so as to provide new ideas for research work in this field.

Aiming at the problems that engine operation data are multi-modal and it is difficult to achieve effective engine life prediction， a parallel branching engine life prediction method was proposed on the basis of multi-modal features integrating potential relationship between images and engine operation time data. Firstly， a sliding window was used to segment the engine operation data， so as to construct sequence samples of engine operation data， and Gramian Angular Field （GAF） was used to convert the constructed sequence samples into images. Then， the sequence samples and images were processed by a Bi-directional Long Short-Term Memory （BiLSTM） network and a Convolutional Neural Network （CNN） to obtain potential relationship features between sensors such as trends and cycles. Finally， Cross-Attention Mechanism （CAM） was introduced to achieve fusion of the two modal features and realize life prediction of the engine. Experimental results on the public C-MAPSS dataset show that the R-squared （R²） of the prediction method is higher than 0.99 and the Root Mean Square Error （RMSE） of the method is less than 1. It can be seen that the method can improve computational efficiency while ensuring the prediction accuracy.

Betweenness centrality is a common metric for evaluating the importance of nodes in a graph. However， the update efficiency of betweenness centrality in large-scale dynamic graphs is not high enough to meet the application requirements. With the development of multi-core technology， algorithm parallelization has become one of the effective ways to solve this problem. Therefore， a Parallel Algorithm of Betweenness centrality for dynamic networks （PAB） was proposed. Firstly， the time cost of redundant point pairs was reduced through operations such as community filtering， equidistant pruning and classification screening. Then， the determinacy of the algorithm was analyzed and processed to realize parallelization. Comparison experiments were conducted on real datasets and synthetic datasets， and the results show that the update efficiency of PAB is 4 times that of the latest batch-iCENTRAL algorithm on average when adding edges. It can be seen that the proposed algorithm can improve the update efficiency of betweenness centrality in dynamic networks effectively.

The outlier detection algorithm based on autoencoder is easy to over-fit on small- and medium-sized datasets， and the traditional outlier detection algorithm based on ensemble learning does not optimize and select the base detectors， resulting in low detection accuracy. Aiming at the above problems， an Ensemble learning and Autoencoder-based Outlier Detection （EAOD） algorithm was proposed. Firstly， the outlier values and outlier label values of the data objects were obtained by randomly changing the connection structure of the autoencoder generate different base detectors. Secondly， local region around the object was constructed according to the Euclidean distance between the data objects calculated by the nearest neighbor algorithm. Finally， based on the similarity between the outlier values and the outlier label values， the base detectors with strong detection ability in the region were selected and combined together， and the object outlier value after combination was used as the final outlier value judged by EAOD algorithm. In the experiments， compared with the AutoEncoder （AE） algorithm， the proposed algorithm has the Area Under receiver operating characteristic Curve （AUC） and Average Precision （AP） scores increased by 8.08 percentage points and 9.17 percentage points respectively on Cardio dataset； compared with the Feature Bagging （FB） ensemble learning algorithm， the proposed algorithm has the detection time cost reduced by 21.33% on Mnist dataset. Experimental results show that the proposed algorithm has good detection performance and real-time performance under unsupervised learning.

As a key module of the question answering system, question classification is also a key factor that restricts the retrieval efficiency of the question answering system. Aiming at the problems of complicated semantic information and large differences of user questions in agricultural question answering system, in order to meet the needs of users to quickly and accurately obtain classification results of common crop disease questions, the question classification model of common crop disease question answering system based on Bidirectional Encoder Representations from Transformers (BERT) was constructed. Firstly, the question dataset was preprocessed. Then, Bidirectional-Long Short Term Memory (Bi-LSTM) self-attention network classification model, Transformer classification model and BERT-based fine-tuning classification model were constructed respectively, and the three models were used to extract information of questions and train question classification model. Finally, the BERT-based fine-tuning classification model was tested and the impact of dataset size on classification results was explored. The experimental results show that, the BERT-based fine-tuning common crop disease question classification model has the classification accuracy, precision, recall, weighted harmonic mean of accuracy and recall higher than those of the Bi-LSTM self-attention network classification model and the Transformer classification model by 2-5 percentage points respectively. On Common Crop Disease Question Dataset (CCDQD), it can obtain the highest accuracy of 92.46%, precision of 92.59%, recall of 91.26%, and weighted harmonic mean of accuracy and recall of 91.92%. The BERT-based fine-tuning classification model has advantages of simple structure, few parameters and fast speed, and can efficiently classify common crop disease questions accurately. So, it can be used as the question classification model for the common crop disease question answering system.

To solve problems such as easily getting trapped in local optimum and slow convergence rate in BareBones Particle Swarm Optimization (BBPSO) algorithm, a dynamic Multi-Subgroup collaboration Barebones Particle Swarm Optimization based on Kernel Fuzzy Clustering (KFC-MSBPSO) was proposed. Based on the standard BBPSO algorithm, firstly, kernel fuzzy clustering method was used to divide the main group into several subgroups, and the subgroups optimized collaboratively to improve the searching efficiency. Then, nonlinear dynamic mutation factor was introduced to control subgroup mutation probabilities according to the number of particles and convergence conditions, the main group was reconstructed by means of particle mutation and the exploration ability was improved. The main group particle absorption strategy and subgroup merge strategy were proposed to strengthen the information exchange between main group and subgroups and enhanced the stability of the algorithm. Finally, the subgroup reconstruction strategy was used to adjust the iterations of subgroup reconstruction by combining the optimal solutions. The results of experiments on six benchmark functions, such as Sphere, show that the accuracy of KFC-MSBPSO algorithm has improved by at least 11.1% compared with classical BBPSO algorithm, Opposition-Based Barebones Particle Swarm Optimization (OBBPSO) algorithm and other improved algorithms. The best mean value in high dimensional space accounts for 83.33% and has a faster convergence rate. This indicates that KFC-MSBPSO algorithm has good search performance and robustness, which can be applied to the optimization of high dimensional complex functions.

Adaboost is an effective classifier combination method,which can improve classification performance of weak learner. Adaboost algorithm was used to resolve facial expression recognition, several combination ways of Adaboost and Principal Component Analysis(PCA) were discussed. Experiment results show that the method has higher classification accuracy than standard PCA.

A texture image segmentation algorithm based on wavelet transform and kd-tree clustering was studied in this paper. Firstly, texture features of an image were extracted using wavelet transform. Secondly, an improved algorithm based on quarter partition was given to smooth the texture feature image. Thirdly, the clustering algorithm using the kd-tree data structure was applied to the texture segmentation, and then a fast texture feature clustering effect was achieved. At last, simulations were performed on the presented algorithm, and the simulation result showed that the presented algorithm has lower segmentation error rate, higher accuracy and better in-time performance.