Research And Implementation Of Multi-pedestrian Tracking Based On Detection

Visual tracking is an important research field of computer vision.In the study of visual tracking,the pedestrian tracking is one of the most important and valuable research topics.Pedestrians' posture is more varied than other targets,and more problems will be encountered when tracking.At the same time,in the actual tracking environment,it is often the case of a number of pedestrians,the multi-pedestrian tracking is more complex than the single target tracking,including the dynamic changes of the target number,the occlusion,the merging and separation between the targets.A tracking method based on detection is often used for multi-target tracking,which first detects the object and location,then carries out the object association,and finally realizes the tracking.In this paper,multi-pedestrian tracking problems under general surveillance scene are studied,focusing on effective detection and tracking of human objects in occlusion,illumination changes,scale changes and number dynamic changes.The main research results are as follows:On solving the pedestrian detection problem,a deep neural network model based on MobileNet structure is proposed.The method utilizes the strong representation power of deep features to improve pedestrian detection accuracy.Meanwhile,it also uses MobileNet as the main network which can greatly decrease the detection time due to fast convolution operations.Experimental results on the public VOC2012 dataset and our own dataset show that the detection method can detect and locate pedestrians in a not only fast but also accurate way.On pedestrian tracking problem,aiming at the problem of particle degradation and particle scarcity in the resampling phase of the classical particle filter tracking algorithm,the wind driven algorithm is proposed to improve it.At the same time,the color histogram feature,HOG feature and LBP feature are combined to enhance the expression ability of pedestrian object,which improved the observation ability of objectsobservation model in particle filter tracking framework.A multi-pedestrian tracking algorithm framework based on detection is proposed.First of all,pedestrian detection algorithm based on deep convolution neural network is used to detect and locate pedestrians accurately.Secondly,the object association between pedestrians is carried out,and multi-pedestrian tracking is divided into single object tracking problem.Finally,the improved particle filter tracking algorithm is used to track,and a solution is proposed to solve the problems of the disappearance of the object,the emergence of the new object,the merging and separation of the object.Based on Windows10 and VS2015,a multi-pedestrian detection and tracking system is designed to realize the automatic detection and tracking of multiple pedestrians in the monitoring scene.