Research Of Object Detection Based On Deep Learning And Target Tracking Based On Kernelized Correlation Filter

The technologies of object detection and target tracking are research hotspots in the field of computer vision.With the rapid development of technology,they have been applied to various fields such as life and military.Although many excellent algorithms have been proposed,the real-time requirement of detection algorithm is high in practical application scenarios,and scale changes,occlusion and other factors will have a great impact on the tracking algorithm.There are still great challenges in target detection and tracking technology.This thesis mainly studies the target detection algorithm based on deep learning and the target tracking algorithm based on kernel correlation filtering,and makes corresponding improvements to the existing problems.The main contents and improvement works of this thesis are as follows:After studying the object detection algorithm based on deep learning,this thesis selects the Yolov4-tiny algorithm as the basic algorithm,and proposes an improved lightweight detection network L?YOLO.The backbone network of the algorithm is composed of CSP structure and DPD structure which first increases the dimension and then decreases the dimension,and the deep separable convolution is used to greatly reduce the amount of calculation and parameters.In the neck part of the network,FPN+PAN structure is used to transfer features from top to bottom and from bottom to top at the same time,which improves the performance of the detection algorithm.And a 52×52 detection layer is added to make the network more adapt to the targets with variable size and improve the detection effect of small targets.In the training phase,the K-means++algorithm is used to cluster the width and height of the real target box of the training set,which can get the initial candidate box that is more consistent with the width to height ratio of the target in the data set,accelerate the network convergence and improve the detection accuracy.In addition,the training data are processed by rotation enhancement and the Mosaic enhancement,so that the network generalization performance is better and the detection accuracy is further improved.Compared with Yolov4-tiny,the storage capacity of L-YOLO algorithm proposed in this thesis is about 1/13,the FPS is increased by 228,and the m AP is increased by 1.72%for specific targets.It can be seen that the improved algorithm has advantages in storage capacity,real-time performance and accuracy.The target tracking algorithm based on kernel correlation filtering can balance tracking accuracy and speed better.Therefore,this thesis improves the KCF tracking algorithm.The principle and shortcomings of the algorithm are studied,and an improved algorithm L?KCF is proposed.First,color features are incorporated into the original algorithm to make the features complementary.In order to solve the problem of fixed size and target occlusion,the improved algorithm introduces outlier detection.The PSR and the SPSR values of the tracking response graph are calculated.When the SPSR value is less than or equal to the threshold,the tracking result of the current frame is considered reliable,and the KCF algorithm is used to track normally.Otherwise,it is considered that the scale change or target occlusion has occurred,and the KCF operation is suspended,and the PSR and PSR_old of the previous frame are recorded for the calculation of SPSR after KCF recovery.The L?YOLO-d algorithm is used for auxiliary detection.Whether the position obtained by the algorithm is in the KCF sampling frame and the confidence degree is used to judge whether the scale changes or the target is occluded.Different processing methods are used until reliable detection target is captured again.The tracking wave gate and KCF model are updated based on the detection results,and the KCF algorithm can resume normal tracking.Compared with the original KCF algorithm,the accuracy of L?KCF algorithm is 6.1%higher,the success rate is 9.9%higher,the robustness is better,and it can better deal with scale changes and target occlusion.