Design And Implementation Of Pedestrian Detection Algorithm Based On Deep Learning

With the continuous development of deep learning technology and computer hardware,the field of computer vision has begun to apply deep learning technology on a large scale to complete tasks such as image classification,object positioning and detection.Pedestrian detection,as an import branch of computer vision,plays an important role in the fields of autonomous driving,human-computer interaction and intelligent monitoring.It is also a technical premise for tasks such as human pose estimation and pedestrian tracking.The task of pedestrian detection is to mark the pedestrian objects contained in the picture in the form of bounding boxes.With the application of deep learning technology,the accuracy of pedestrian detection has been greatly improved,but there are still many problems.First,pedestrian occlusion is still a challenging problem in pedestrian detection task.Second,deep learning detection algorithm are difficult to deploy and apply on embedded devices or mobile terminals with limited computing power.In response to the above problems,the main content and results of this article are as follows:(1)Implemented the improvement of the CSP algorithm for the pedestrian occlusion problem.The CSP algorithm is improved by using the idea of pedestrian head features helping the overall pedestrian detection.In the feature extraction module,the extraction of pedestrian head features is added to enhance the overall features of pedestrians,and to make the algorithm have better multi-scale target detection capabilities without increasing the amount of calculation,the pyramid interpolation module(PIM)is used for feature fusion.To enhance the expression ability of pedestrian head features and overall pedestrian features,separate detection of pedestrian heads and overall pedestrians is added in the detection head module,and multi-task joint training is adopted.The improved algorithm has better detection accuracy for pedestrian occlusion objects and multi-scale objects.(2)Implemented the accelerated optimization of the algorithm.The algorithm is optimized by using the lightweight backbone network MobileNetv2 and the TensorRT acceleration library.The former reduces the number of parameters of the algorithm,the latter optimizes the network structure,and accelerates the detection speed of the algorithm through low-precision representation of parameters.Real-time detection can be achieved on the embedded device Nvidia Jetson Nano.(3)Designed and implemented a pedestrian detection system.The system uses the pedestrian detection algorithm implemented and optimized in this article to process the video stream in the camera in real time,and send the detected pedestrian object to the user's Android client in the form of a picture.The user can also view the monitored real-time picture or Video recording.