Research And Implementation Of Pedestrian Detection And Tracking Based On Embedded

Computer Vision is an important field dedicated to the research of computer graphics.Pedestrian detection is currently a hot topic in computer vision research.It has huge economic value and potentiality in the fields of social public safety,business activity analysis,military defense security,and smart transportation.In this paper,in terms of pedestrian detection,based on the YOLO-v3 algorithm,the detection principle and network training process of the algorithm are studied in detail,and a model optimization method based on adaptive spatial feature pyramid fusion is proposed.The RFB module effectively integrates deep-level features and improves the recognition rate of pedestrian detection models.Then integrate the pedestrian images of the public data sets voc,coco,sysu and prw,construct a pedestrian data set,and use the pytorch framework to complete the data training.The final model test set mAP reaches 78.96%.The pedestrian image selection through the network has a good test result and proves that the model is generalized excellent ability.In terms of pedestrian multi-target tracking,the deep sort algorithm for multi-object tracking is used,Then,analyze the specific algorithm principle of deep sort and propose some optimization strategies.Finally,based on the previously obtained YOLO-v3 optimization model,the combination of detection and tracking,Complete the integrated design and test it on the MOT16 data set and the actual driving recorder.The performance is good,and the purpose of high precision and real-time performance is met.According to the YOLOv3 pedestrian detection and tracking optimization algorithm proposed in this paper,a specific implementation based on the embedded board jetson-nano is made.First,the reasons for choosing NVIDIA jetson-nano as the operating platform of this system are described.Nano,as NVIDIA's own embedded product,has a cuda core to facilitate the transplantation of GPU algorithms and speed up R&D efficiency.Then based on the acceleration characteristics of TensorRT,the pedestrian detection model used in this paper is optimized and accelerated,and some model pruning methods are proposed,and model pruning is done on this basis.For the current construction data set,after channel pruning of YOLOv3,the model parameters and model size are reduced by 80%,FLOPs are reduced by 70%,the speed of forward inference can reach 200%,and the mAP can be kept basically unchanged.Finally,the current mature solution,Pytorch model-> ONNX-> TensorRT model,is adopted to successfully realize the model optimization acceleration of the embedded side.Finally,the simplified line of the line is considered,and the pedestrian monitoring system is built and tested by the method of client-side multi-concurrency mechanism ImageZMQ wireless transmission video and server-side reception,and the system can process multiple channels of video at the same time,by adjusting the frame size,Speed up processing.The processing speed is about 20 fps,which is about 2 times higher than the original model,which proves that the system is stable and reliable,and can be used to detect and track pedestrians and pedestrian analysis statistics in real time.