A Robust Tracking Method For Video Target Under The Condition Of Occlusion And Illumination Change

With the deepening of artificial intelligence research and the rapid development of computer technology,video target tracking technology is widely applied in many fields,such as public security,driverless cars,intelligent interaction,military attack and many other fields,forming the bottom of many intelligent systems.Because of interference factor such as occlusion or illumination changes,it often leads to low tracking accuracy or even tracking failure,which directly affects the normal operation of intelligent system.Based on kernel correlation filter tracking,we study the occlusion and illumination problems from three aspects: multi-features fusion modeling,anti-noise tracking algorithm improvement and template difference updating.1.The multi feature adaptive fusion modeling is studied in the kernel correlation filter framework.A kernel correlation filter tracking method based on multi-features adaptive fusion is proposed to solve the problem that feature used in the kernel correlation filter tracking algorithm is single and has poor robustness to illumination change.Firstly,make full use of the chrominance components of HSV color model in light sensitivity,and extract the target color component histogram of gradient,and retain the spatial information of hue histogram feature to kernel weighted method,and then use the nuclear related filter framework set adaptive fusion coefficient of mean square loss function in the process of training features.Finally,using the multichannel processing characteristics of Gauss kernel function for decision level fusion of multi-features weighted.2.The tracking framework of joint estimation of multi block location cues based on kernel correlation filter is studied and constructed.Aiming at the problem that kernel correlation filter algorithm is not strong enough in noise immunity and easy to be interfered by external factors such as occlusion and optical noise,the tracking accuracy is reduced.A kernel correlation filter tracking method based on multi-blocks joint estimation is proposed.According to geometric characteristics of the initial frame tracking frame of target adaptive block,and the use of kernel correlation filter algorithm of sub block independent tracking combined with confidence map,then the location and scale of above frame target as apriori information on the search area at the same time sampling,the sampling frame confidence graph weight density as the observed value,realization on the overall optimal candidate target location and scale are estimated using particle filter algorithm.3.The method of template differentiation updating was studied.According to introducing model easy to shelter and optical noise error correlation filter algorithm the existing template updating process and lead to tracking drift problem,based on the kernel correlation filter framework,multi block joint estimation is proposed.In a reverse projection sub block occlusion detection judgment method based on template.Reverse projection to the region first confidence lower blocks located in the search area,the search area and then statistical foreground pixels and background pixels,to determine the probability of sub positioning area belongs to target and background,so as to realize the difference of template update.