Research On Adaptive Target Tracking Algorithm

As one of the main branches of computer visualization technology,target tracking technology has played a significant role in the application fields of automatic driving,intelligent monitoring,robot and so on.In recent years,with the deepening of the research on correlation filtering and deep learning target tracking algorithms,they have made great breakthroughs in the instantaneity and accuracy of tracking respectively.However,in the actual scene,target tracking will face various uncertain factors such as target scale transformation,deformation and so on.Therefore,this paper discusses and studies the development of existing target tracking technology,and puts forward some adaptive solutions based on existing tracking algorithms.The main contents are as follows:(1)The performances of KCF,CSK and dsst tracking algorithms based on correlation filtering are compared through experiments.In the experiment,the three algorithms are put into the otb-100 data set to track the target in a variety of scenes such as scale transformation,occlusion and complex background,show the specific tracking effects of the three tracking algorithms in the actual scene,and evaluate them with the relevant indicators of target tracking.According to the analysis of experimental results,because the target frame size of KCF algorithm cannot be changed,it has poor tracking effect in the face of scale transformation scenes,but it can also show good robustness in the face of some complex scenes.(2)A scale adaptive method based on KCF tracking algorithm is proposed.This method uses the existing inaccurate tracking information to set the target range box,subtracts the images of the front and rear frames in the range box,eliminates the irrelevant background information,and obtains the binary image in the range box.Filter the pixels in the binary image,determine the size and position of the target,and replace the scale and position of the target frame in the next frame,so that the tracking matching template can be updated,so as to track the object more effectively.This method is used for experimental test,and the results are compared with those of KCF method.This method can reset the more accurate target frame when the object motion has a large-scale transformation,so as to update the matching template and track the target for a long time.(3)A tracking method based on Yolo detection is proposed.This method uses the yolov5 s model in the deep learning algorithm to detect the initial frame of the video sequence,identify various types of objects in the image,and screen out the targets to be tracked.After the tracking target is determined,the detected image is scaled and only the target is displayed.Next,the tracking of the target object can be realized only by frame by frame detection.Because the reduced display frame reduces the calculation range in the detection process,it improves the calculation speed and meets the real-time requirements of tracking.At the same time,because the algorithm adopts the operation of detecting the target in the tracking process,it can directly identify and label the target,so this method has the potential of retraining after the target is lost.