Research On Tracking Algorithm Of Scale-Adaptive Correlation Filter Based On Object Deep Feature

As one of the popular research directions of computer vision,object tracking technology has been widely used in intelligent security,intelligent transportation,human-computer interaction,military and other fields.However,current object algorithms still face many difficulties,such as insufficient feature expression capabilities,scale changes,object loss,etc.This article focuses on these difficulties and makes some contributions under the framework of correlation filter tracking algorithms.The main research content is divided into the following two aspects:1)An improved object tracking algorithm based on ECO is proposed.The following work has been done on the basis of ECO: firstly,a channel screening method based on the object energy ratio is proposed to screen features channels according to the energy proportion of the object in the sample area and finally reduce the dimensionality of the deep features;secondly,a weighted fusion method is proposed to fuse the correlation response maps of different convolutional layer features with different weight;thirdly,a fast multi-scale detection method with variable aspect ratio is proposed,which divides object positioning and scale detection into two independent tasks,and uses samples with different scales and aspect ratios to train scale filters.It can detect not only the change of the scale but also the change of the aspect ratio of the object bounding box.The experimental results on the data set OTB100 show that the distance accuracy and overlap success rate of the improved algorithm are only 0.33% and 1.31% lower than ECO,but the tracking speed is2.90 times higher than ECO.On the basis of keeping the distance accuracy and overlap success rate close to ECO,the tracking speed has been greatly improved.2)An object tracking algorithm combining DCF and YOLOv3 is proposed.The tracking task is divided into two stages: rough positioning and precise positioning.In the rough positioning stage,tracking algorithm DCF is used to predict the rough position of the object.In the precise positioning stage,all objects of the same kind as the object are detected in the rough position area by using the detection algorithm YOLOv3;based on the detection results,interference warning of the same type objects is proposed,and the object is judged as having no or having same type interference status: for the state of having no interference of the same type,locate the object from the background of the same type according to the object category;for the state of having interference of the same type,a CN discriminant model is proposed,and the object is located from the same object based on the Color Names(CN)feature.For the problem of object loss,a object re-search method is proposed,which re-searches the object according to the object type and CN feature in the search area.The YOLOv3 multi-scale detection method is used to solve the problem of scale change.In the data set OTB100,57 color video sequences are selected for testing.The experimental results show that the distance accuracy and overlap success rate of this algorithm are 86.81% and 83.35%,respectively,and it has good tracking performance.