Research On Video Object Detection And Tracking Based On Hybrid System

Visual object detection and tracking is the important content in the field of computer vision.With the increasing computational power,the emergency of low-cost high-quality video capture device,and the increasing demand for intelligent video surveillance system,object detetion and tracking algorithms are extensively researched.To take advantage of the computer instead of human engaging in motion perception,scenes understanding,etc,many difficulties should be overcomed for object detection and tracking method,such as illumination variations of the monitoring scenario,pose variations of tracked object,the spatio-temporal uncertainty of object disappearing and reappearing and so on.This thesis realizes the pedestrian tracking in presence of missed detection and full occlusion based on hybrid system framework by establishing different levels of video representation model,design logic switching rules to solve complex scenes moving target detection and long-term tracking.The innovations of this thesis are as follows:1.An object detection algorithm is proposed based on hybrid system.The proposed method uses the color difference histogram video segmentation method and codebook background modeling method to construct the model set.A set of logic transition rules are designed to make full use of both the color difference histogram algorithm with low complexity and codebook background modeling algorithm with complex background modeling ability,meanwhile,those rules can effectively avoid the shortage of two algorithms in dealing with camera shake and illumination variations or real-time target detection.The experimental results show that the proposed algorithm can overcome the dramatic illumination changes and camera shake effect on video segmentation,and realize a fast and robust moving target detection.2.For the multiple-object tracking with birth target,a track initiation method and the Gaussian mixture probability hypothesis density(Gaussian mixture probability hypothesis density,GM-PHD)filter method are ultilized to construct a hybrid model.The transistion rules are designed to deliver the GM-PHD filter method with birth target and properly allocate measurements to the track initiation method so as to achieve birth target identifying and all targets tracking with less calculation.Experimental results show that the proposed method can obtain a good tracking performance and real-time in case of unknown and time-varying number of targets.3.For object tracking in the presence of full occlusion,tracking model and re-identification model are used to construct the hybrid model set.A set of reliable model transition rules are designed to ensure feature learning in tracking stage,so as to achieve re-identification the tracked pedestrian using the re-identification model from the detection result after the tracked pedestrian is fully occluded,and reseting the state of the tracked pedestrian after the tracked pedestrian is identified.Experimental results show that the proposed method has significant advantages for object tracking in the presence of full occlusion.4.A multiple-object tracking method based on hybrid system is proposed in the presence of miss detection.A Gaussian component that fused both appearance and state information together with the two terms in GM-PHD update components are used to construct the hybrid model set.Whether the tracked target is miss detected that is judged by the set transition rules.Multiple object tracking in the presence of miss detection is achieved by using the optimum model each time.Experimental results demonstrate that the outperformace of the proposed method for object tracking in the presence of miss detection.