The Object Detection Method For Pedestrian Video Based On YOLOv3

In recent years,with the development of Smart City and the application of video monitoring,pedestrian detection is playing an increasingly important role in the field of computer vision.Pedestrian detection is a technique to identify and locate pedestrian target in the image or video.Pedestrian detection has broad application prospects both in the military and civilian field.Pedestrian detection is always a challenge in the field of computer vision.Because the existing target detection algorithms are still inaccurate in target positioning,the problems in miss detection and detection accuracy needs to be improved.Based on the most advanced target detection algorithm YOLOv3,the improvements in image partition size,k-means and detection module are proposed in this paper.On the other hand,most data sets of pedestrian detection come from monitoring or aerial photography video in practical engineering.However,there is no labeling software for video data on the market.Therefore,in this paper,a labeling software for video is designed and further improved according to the characteristics of pedestrian targets.The main work and innovations of this paper are as follows:1.Pedestrian targets are usually small and dense.There are some problems in using YOLOv3 for pedestrian detection,such as poor detection accuracy,inaccurate target positioning,frequent miss detection and so on.Based on the above problem,three improvements for YOLOv3 are presented in this paper.Firstly,the comparison test are conducted.When the precision and speed of detection are took into account,it is proposed to use 1010? as the image partition size.Secondly,a modification method in initialization of candidate box is proposed for k-means clustering algorithm.In this method,the data filters are added before clustering in order to make the size of candidate box more suitable for pedestrian characteristics.Finally,a construction of 2-way dense layer is introduced into the detection module to improve the recognition ability of network for features with various sizes.To confirm the algorithm's generalization,the comparison tests of the modified YOLOv3 and original algorithm are conducted on the VOC data sets.2.Most data sets of pedestrian detection come from video.Facing to the lack of available software to label video data on the market,a labeling software for video is developed.Based on the characteristics of pedestrian video,the key frame extraction function based on inter-frame difference and the auxiliary function based on edge detection are added.The optimized software improves the ability of key frame extraction and detection accuracy.Finally,the software is used to label the pedestrian video captured around Xidian University and the results are shown.In the detection accuracy aspect,the test results are below: The improvement of image partition size increases the m AP of YOLOv3 on pedestrian data set from 86.6 to 87.1.The improvement of candidate box size initialization further increases the m AP to 87.3,and the improvement of detection module makes the m AP reach 89.4.Compared with the original YOLOv3.The improvements in this paper result in a total detection accuracy increase of2.8%.In terms of detection speed,the original YOLOv3 detection speed on the pedestrian data set was 52.59 fps,while the improved YOLOv3 detection speed on the same data set was 47.75 fps,which was reduced by 9.2%.