Research Of Pedestrian Detection System Based On Deep Learning

Pedestrian detection has great value of research and extensive application prospect,since deep learning technology has achieved great success in target detection in recent years and pedestrian detection is a specific task of target detection,the thesis selects the YOLOV3 algorithm with better performance from the target detection algorithms to conduct research on pedestrian detection system based on deep learning,the main contents are summarized as follows:The YOLOV3-MPL pedestrian detection network is proposed.The technical principle of YOLOV3 algorithm is analyzed in detail,on the basis of which the anchor mechanism is improved for widely distributed size and relatively fixed aspect ratio of pedestrian and the MPL network element is added to the original network structure to highlight the useful features of pedestrians.Then the main ideas involved in the model is discussed,the multi-layer feature acquisition and feature procession technology in solving wildly distributed pedestrian size are analyzed,some technical details during the model training process are introduced.Experimental results on the dataset indicate that the added MPL network unit,improved anchor mechanism,and multi-layer feature acquisition technique all can improve detection accuracy,thus verifying the rationality and effectiveness of the network structure.The pedestrian dataset is made.Some screenshots of pedestrian pictures are took from the video saved by driving recorder to make the dataset needed in experiment by using dataset making software,the self-made dataset consists of 1021 pedestrian images,in which 700 belong to training set and the rest belong to testing set.The YOLOV3-MPL model is successfully transplanted into the embedded platform NVIDIA TX2.The pedestrian detection system based on deep learning is successfully built on embedded platform through tailoring the network partly.The test results on the embedded platform show that although the tailored model has a reduction of about 40% in terms of average accuracy,the average detection rate has become four times the original,reaching 16 frames per second.The work of this thesis verifies the possibility of real-time pedestrian detection system based on deep learning on embedded platform,as a result of which has certain theoretical reference value for realizing advanced driver assistance system in the future.