Multi-Camera Pedestrian Detection And Tracking System Based On ARM Embedded Platform

Wireless multimedia sensor networks have great application prospects in the field of environmental monitoring and disaster relief.According to the actual demand of multimedia sensor network,this thesis studies the omnidirectional sensing nodes and pedestrian detection and tracking algorithm.Based on the ARM embedded platform design WMSN omnidirectional sensing nodes,using six-channel wide-angle cameras which are placed in a coaxial way to form an omnidirectional monitoring system,which realizes multi-channel information acquisition and pedestrian detection,and continuous tracking of interest target,has a strong application value.In order to meet the real-time nature of embedded platform pedestrian detection algorithm,this thesis analyzes the basic framework of pedestrian detection,aiming at the problem of poor real-time performance of traditional sliding window-based methods,proposes a recommended window strategy based on foreground detection.Because the camera is placed still,this thesis first uses Vibe algorithm to detect moving objects.Based on the detection results,Hog features are extracted and trained by trained SVM classifier to achieve pedestrian recognition.The Vibe algorithm has a small amount of calculation and can produce a complete target area,greatly reducing the number of proposed windows.It greatly improves the real-time performance of the algorithm under the premise of the accuracy.To achieve continuous tracking within the field of view,a stable single camera target tracking algorithm is firstly needed.One notable difference of the tracking algorithm over the general target detection algorithm is that the target being tracked has a spatio-temporal context.The full use of temporal and spatial context information can effectively improve the robustness of the tracking algorithm,therefore a long-term target tracking algorithm is proposed in this thesis.Based on the temporal and spatial context information,we achieve long occlusion decision in target tracking process.The occlusion decision module is used to combine the STC tracker and the online training cascading detectors to realize the target re-recognition and follow-up tracking after long-term occlusion.Through experiments,the long-term tracking algorithm proposed in this thesis has good performance in real-time and robustness.In order to realize the omnidirectional perception,the system uses multiple cameras to form an omnidirectional field of view,which involves the problem of multi-camera object handover with overlapping viewpoints.In this thesis,according to the angle of view,brightness and distortion introduced by wide-angle camera in multi-camera,an algorithm combining visual field dividing line and target matching to achieve accurate target handover is proposed:First,we use the wide-angle camera vision field generation method with overlapping views to generate the boundary line of vision,and estimate the timing of the transfer of the target,and make a rough estimate of the possible target area.On this basis,the target matching is achieved by a cascade method based on the matching of color attribute features and spatial color features,and achieves efficient target handover method by cascading the target matching algorithms with different accuracy and real-time.Finally,this thesis designs software and hardware system,builds a high-performance hardware platform based on PCIE video capture card and i.MX6Q processor,and carries out the algorithm embedded platform transplanting.Due to the limited computational ability of the embedded platform,this thesis optimizes the algorithm based on the hardware and software environment.Finally,the algorithm is verified and analyzed based on the omnidirectional sensing node prototype system.