| Pedestrian detection and re-identification,the challenging research topic in the field of com-puter vision and important steps in the construction of intelligent surveillance systems and smart cities,have attracted widespread attention and research from academic and engineering circles for their important theoretical significance and practicality.However,traditional detectors are hard to satisfy the requirements of efficiency and accuracy simultaneously,efficient detectors are easy to miss targets,and high-precision detectors are hard to achieve real-time operations.For pedes-trian re-identification,there remains a large gap with research and application,and the accuracy could be improved.This thesis mainly focuses on theoretical research and extensive experimen-tal verification around improving detection speed and re-identification accuracy.The algorithm is packaged in software and embedded in Jetson TX2 to build a high-performance intelligent surveil-lance camera equipment,which breaks the barrier between academic and engineering.The main innovations of this thesis are as follows:1.We construct three lightweight convolution units based on depth-wise separable convolution and using channel shuffling,channel splitting,and identity mapping.The lightweight units proposed maintain the performance of the model while reducing the number of parameters and calculations by nearly 7 times compared to ordinary convolution.Training and eval-uation of classification tasks on ImageNet demonstrate the performance advantage of the proposed method over other lightweight units.2.For the construction of a lightweight pedestrian detection network,the idea of YOLOv3 is used,and the backbone network is constructed by using the three types of lightweight con-volution units proposed.Subsequent networks widely use group convolutions,which re-duces the model to the original parameter in terms of YOLOv3 15%,the speed is increased by about 1.7 times,and the network calculation can reach 200 FPS.The selection of the number of clusters based on the average distance and the silhouette coefficient for K-means clustering,and the dimension of the anchor boxes clustering by the face and the human re-spectively and combining them,which reduces the error of the predicted coordinates by 1%.The margin cross-entropy loss is proposed to guide the confidence training,which is 4.6%mAP higher than the traditional cross-entropy training.Comprehensive experimental results on three challenging pedestrian benchmarks demonstrate the effectiveness of the proposed approach.3.For the pedestrian re-identification network,the structure of parallel extraction of local and global features was designed.ArcFace was introduced for classification learning.The pedestrian detection model was combined to improve the quality of input pictures,which reduces useless background information,avoids the process of human body alignment,and makes it possible for raw video frames to be used directly as input to a re-identification network.A visibility scoring subnet is proposed for weighted distance calculation in con-junction with local feature vectors to solve the problem of incomplete or highly occluded recognition.The model is evaluated on two public datasets and achieves the highest 86.0%mAP and the best 94.1%rank-1 accuracy rate on the market-1501 dataset.4.The proposed models are packaged into a software written in PyQt,which have the advan-tages of cross-platform operation,high compatibility,and beautifully designed interfaces.The intelligent surveillance camera is built with Jetson TX2 and our software,which elimi-nates dependency on the traditional video server and achieves high-performance and high-quality edge computing. |
PDF Full Text Request