Research On Pedestrian Detection Algorithm Based On Embedded Platform

Pedestrian detection is one of the hot research fields in computer vision,and it has important applications in security monitoring,auto autonomous driving,intelligent robot and other scenarios.In recent years,pedestrian detection technology based on deep learning has developed rapidly,with great improvement both in accuracy and speed.In addition,with the development of edge computing,real-time pedestrian detection based on embedded platform has become a research hotspot.Aiming at the pedestrian detection algorithm on the embedded platform,this thesis mainly carries out the following research work:First of all,due to the limitation of computing resources on the embedded terminal,this thesis selects the YOLOV3-Tiny model for in-depth study by checking balance between the detection accuracy and computing speed,and designs a high-performance pedestrian detection network suitable for the embedded system.In terms of Network structure,Cross Stage Partial Network(CSPNET)and Spatial Pyramid Pooling(SPP)structures,are used to strengthen the feature extraction ability of YOLOV3-Tiny Network and enrich the expression of feature map.Considering that the size of the target in pedestrian detection task varied greatly and there were many small targets,the prediction branch is increased to three,and a bidirectional Feature Pyramid network(FPN)is designed to integrate more features.In order to improve the positioning ability of the prediction frame and accelerate the convergence of the model,Distance Intersection over Union(DIOU)loss is also used in this thesis.The experimental results show that the enhanced YOLOv3-tiny model has a better detection effect.The average precision(AP)on the Caltech,KITTI and INRIAPerson data sets is 78.96%,80.68% and90.30%,respectively.Therefore,pedestrian detection tasks in complex backgrounds can be completed in a high speed and accurcy.The model is then deployed on a Nvidia Jetson Nano 2GB embedded platform.According to the computing characteristics of the platform,the model is optimized from two aspects: first,the deep separable convolution and linear bottleneck structure are introduced to simplify the network structure and reduce the amount of calculation;secondly,using the Tensorrt model inference acceleration library,the framework of pedestrian detection acceleration reasoning is designed.The experimental results show that the combination of the lightweight operation and Tensor RT can make the model achieve greater speed improvement with less loss of precision.Finally,the accuracy,real-time performance and stability of the proposed pedestrian detection algorithm are verified on the embedded Nano 2GB platform.Deploy the model and acceleration framework designed in this thesis to Nano 2GB,collect data with IMX219 camera,and use V4L2 and GStreamer to codec and decode video to build an efficient embedded pedestrian detection system.The experimental results show that the pedestrian detection system designed based on the embedded platform in this thesis can detect the pedestrian target quickly and accurately.When the video resolution is 640×480,it can work stably at the frame rate of25 fps.