Research On Video Target Detection And Tracking Algorithm Based On SNMM And MDP

Target detection and tracking has always been the focus of research in the field of computer vision,and is widely appied in unmanned driving and video surveillance.Due to factors such as changes in lighting in the video,shadows of targets,occlusion between targets,disappearance and rebirth of targets,target detection and tracking has become a challenging research topic.In this thesis,we focus on the target detection based on skew normal mixed model(SNMM)and the multi-target tracking based on the Markov decision process(MDP).The specific research work is as follows.(1)An improved SNMM is proposed to solve the problem of detection failure caused by light changes in videos.The position parameter and skewness parameter in SNMM can well reflect the change of lighting.Based on this,a light change detection criterion including position parameters and skewness is established.When light changes are detected,the video frame is processed by background subtraction,which greatly improves the accuracy of target detection.Compared with Gaussian mixture model(GMM),SNMM and background difference method,the improved SNMM detection target has the highest performance evaluation.(2)The video multi-target tracking model based on Markov decision process is divided into 4 states:Active,Tracked,Lost,and Inactive.The quality of the transition strategy between states determines the accuracy and speed of target tracking.In order to improve the speed of target tracking,a fast MDP(Fast_MDP)multi-tracking algorithm with a fast matching strategy,which uses the number of detected targets to judge the state of the target and uses a scoring matrix to match the target,is proposed between active and lost states.The experiment results show that the tracking speed of Fast_MDP algorithm has been greatly improved.(3)In order to solve the problem of Lost target entering Tracked state again,a target matching tracking algorithm based on image feature similarity is proposed.The image features employ the HSV color histogram and LBP(Local Binary Pattern)histogram.According to the cosine similarity between features and the Jaccard coefficient between targets,a multi-feature similarity matching criterion is established,which solves the problem of tracking the lost target again.The experimental results verify the feasibility of the proposed algorithm.