Research On Target Tracking Method Based On Correlation Filtering In Road Traffic Scene

In recent years,the traffic situation has become increasingly severe.In order to ensure the safety of pedestrians and vehicles,target tracking in road traffic scenes has quickly become a research hotspot in the field of computer vision.Correlation filtering,as one of the basic methods of target tracking,is favored by researchers for its fast tracking speed.However,the real traffic scenes are complex and changeable.Many problems such as fast motion,occlusion,and scale change seriously affect the tracking accuracy of the correlation filtering.Therefore,starting from the above problems,this thesis improved the correlation filtering tracking method.The main research results are as follows:(1)Aiming at the changeable position of a fast-moving target and obvious boundary effect,a position prediction method based on Bezier curve is proposed.Firstly,uses the Bezier curve to fit the historical position of the target.Secondly,predicts the target position of the subsequent frame according to the motion trend obtained from the fitting.Finally,the prediction interval adjustment strategy is introduced,and the prediction interval parameters are dynamically adjusted by evaluating the prediction effect of historical frames,to reduce the unnecessary prediction frequency and improve the accuracy of position prediction.(2)Aiming at the problem that the existing scale estimation methods can only estimate the target scale according to the fixed aspect ratio,an adaptive aspect ratio scale estimation method is proposed.Firstly,this method uses fuzzy clustering to delimit a reasonable block area.Secondly,uses the block tracking idea to determine the location of the sub-block.Finally,estimates the actual scale of the target through the joint centroid scale vector to realize the adaptive aspect ratio scale estimation.(3)Aiming at the problems that the existing correlation filter tracking methods are not accurate enough,the filter template update rate is fixed and easy to be polluted,a multi-method fusion correlation filter tracking method based on adaptive learning rate update is proposed.Firstly,the saliency detection method is used to extract the saliency area.Secondly,the saliency value of all pixels in the area is normalized,and then the learning rate is dynamically adjusted according to the saliency value of the target center position.Finally,the three proposed methods are merged into the correlation filtering framework using depth features to form a new correlation filtering tracking method.The tracking method proposed in this thesis is experimented on the OTB120 data set with traffic video sequence added.The experimental results show that in the road traffic scene,the tracking method proposed in this thesis has higher accuracy and better robustness than the existing tracking methods of the same type.