Research On Target Tracking Algorithm Based On Discriminant Model

As one of the important research areas in computer vision,target tracking refers to maintaining proper attention to the target required by a task in a video and obtaining information about its position,scale,etc.And provide information base for more advanced tasks.Visual single target tracking algorithm is a process of obtaining information on a single target of different types in a continuous video.No matter what the type of the target is,tracking can be achieved without prior knowledge of the target information as input.In recent years,discrimination-based target tracking algorithms have been developing and progressing continuously.However,when encountering complex and difficult environments,the tracker is still prone to tracking failure and drift.In this paper,from the two perspectives of correlation filtering and full convolution twin neural network,from the perspective of feature extraction occlusion discrimination and model updating,the properties of target tracking algorithm facing different challenges is improved and tuning.The main research contents and algorithm improvement of this paper are as follows:(1)Based on the correlation filter tracking algorithm,a dynamic adaptive weighted fusion of HOG features,color features,and grayscale features is performed to complete the position estimation of the position filter;Then,a scale filter is constructed in the central position of target tracking to improve the performance of the algorithm on scale;At the same time,a new occlusion discrimination strategy is proposed to selectively update the tracking model and provide multi-scale search area for different objects.And using the improved algorithm,the object tracking system and the operation interface of selectable objects are designed,which can choose the objects in the video and complete the tracking.(2)Channel attention and space attention are used to sift and weight the features in the process of feature extraction to higher the precision of the siamese fully-convolutional tracking algorithm.In this way,useful feature information can be weighted with a higher weight,so as to extract more representative feature information effectively.The establishment of the tracking model is optimized to enhance the robustness of the tracker and improve the tracking performance of the algorithm in complex scenes.(3)The improved algorithm is tested on a target tracking benchmark test set and compared with other algorithms on relevant platforms.The experimental outcome indicate that the overall performance of the improved algorithm is enhanced,and it has more advantages under the challenges such as deformation,occlusion and low resolution.