| Pedestrian re-identification is a key technology in intelligent surveillance systems,which can automatically detect and recognize pedestrians in monitoring scenes.It has high value in research and engineering application.However,the existing cloud-based pedestrian re-identification algorithms have limitations such as high video data transmission between the edge and the cloud,and heavy computational load on the cloud.In recent years,with the emergence of highperformance embedded hardware and more lightweight net,it has become possible to achieve efficient pedestrian re-identification on embedded platforms.This thesis focuses on the research of pedestrian re-identification technology and its implementation on an embedded platform,with the following main contributions:(1)The fusion of pedestrian detection algorithms and pedestrian re-identification algorithms is studied.The serial and end-to-end pedestrian re-identification methods are introduced.And the overall structures and the pros and cons of the two methods are explained.Then the serial pedestrian re-identification method is selected for addressing and implementation on the embedded platform.(2)The pedestrian detection algorithm based on YOLOx is studied and optimized.The lightweight convolutional neural network Mobile Net V2 is applied to build the backbone network.And the convolutional process of YOLOx is optimized by using depth-wise separable convolution units to reduce the ’model’s parameters and computational complexity.The channel attention mechanism is introduced into the backbone network to enhance the capability of feature extraction.An enhanced feature extraction network is built by using the weighted bi-directional feature pyramid to improve the feature fusion capability.The algorithm performance on the Wider Person dataset is validated,achieving an m AP precision of 65.85%.(3)The lightweight pedestrian re-identification algorithm is studied and optimized.The selection of the backbone networks is Analyzed.And then Mobile Net V2,which has fewer parameters and computations,in addition performs well,is chosen as the backbone network.The feature extraction capability is enhanced by using the coordinate attention mechanism.A multi-task learning branch structure is designed by the idea of combining the representation learning with the metric learning,and the multiple loss functions are united to improve the accuracy and the generalization ability of the model.The algorithm performance is validated on the Market1501 dataset,achieving an m AP precision of 69.2% and Rank-1 accuracy of 86.2%.(4)The pedestrian re-identification function is designed and implemented based on the embedded hardware platform RK3399 pro.This model is transformed,quantized,evaluated,and deployed on the embedded platform.The pedestrian detection module and re-identification module are designed and accomplishes the application in the pedestrian re-identification procedure.To verify the accuracy of the pedestrian re-identification function,it is tested with 122 panoramic images containing 66 persons’ identities,achieving an accuracy of 80.25% on the embedded platform,with a processing speed of 5.88 frames per second.This thesis implements the pedestrian re-identification functionality on an embedded platform,breaking the dependency on cloud servers and demonstrating high application value. |
