Research On Target Detection And Tracking Based On YOLO And Correlation Filtering Algorithm

Moving target detection and tracking is an important research content in the field of computer vision.Based on YOLO and related filtering methods,this paper studies tracking methods for small target detection and target in complex situations such as occlusion,deformation,scale transformation,and environmental lighting changes.The main research contents include:Aiming at the problem that the target detection accuracy of the YOLO algorithm decreases when there are small targets in the environment,an improved YOLO algorithm is proposed.The algorithm proposes an improved calculation method for the loss function on the traditional YOLO algorithm from the two aspects of coordinate positioning and target category determination.The former proposes a calculation method for the loss function of the bounding box width and height based on relative errors.It can also pay enough attention to improve the shortcomings of YOLO algorithm's low detection rate for small targets;the latter proposes a loss function calculation method based on intuitionistic fuzzy numbers for target category prediction,using intuitionistic fuzzy numbers to replace the probability in the loss function The feature information obtained by the convolutional network can be fully used,which effectively improves the accuracy of target detection.The experimental results show the rationality and effectiveness of the proposed algorithm.Aiming at the problem of target tracking performance degradation when there is target occlusion and illumination changes in the environment,an improved KCF tracking algorithm that incorporates motion information detection is proposed.Firstly,using motion information,based on optical flow method and inter-frame difference method to predict the possible areas of the target,so as to reduce the search range and improve the speed of the algorithm.Secondly,based on the traditional KCF algorithm,according to the nonlinear relationship between the learning rate and the peak value of the response graph in the model update strategy,a parabolic learning rate curve is constructed to implement the adaptive update of the model.The tracking accuracy of the algorithm is guaranteed,and the experimental results of related video tracking verify the effectiveness of the improvement.Aiming at the complex situations such as scale change and deformation of the target,a scale change KCF tracking algorithm based on Radon transform is proposed.First,we use Radon transform's insensitivity to noise and moment translation invariance,and then use the proposed online model update strategy to construct a scale filter and position filter based on Radon transform to achieve the scale estimation of the target frame and related video tracking.Experimental results verify the effectiveness of the proposed method.