Research Of Object Tracking Based On Deep Learning Object Detection And Kernel-correlation Filter

Object detection and tracking is an important part in computer vision field.The development of target tracking ranges from generation-based tracking to discriminant tracking to deep learning tracking.But because of the complexity of target tracking,there is currently no effective algorithm to deal with all scenarios.The rise of artificial intelligence has led the development of deep learning,and the accuracy of detection model in deep learning has gradually increased,which has brought dawn to many challenges in multi-target tracking field.The application of deep learning in multi-target tracking is divided into two directions: one direction is that using the deep network completely learns the target tracking,and the other one is that tracking algorithm combines the deep neural network and conventional tracking algorithm.This paper studies the second direction.In addition,the deep neural network generally is a large model,and the parameters and computing power are relatively high.In practical applications,most devices do not have these computing resources,so the compression and optimization techniques of the deep network without loss of precision are also engineering applications.The main technical research on object tracking in this paper is as follows:(1)Based on the Kernel Correlation Filtering tracking Algorithm(KCF),propose a improving MACF algorithm.Aiming at the scale transformation of the target,based on the anchor mechanism in deep learning,a scale scaling factor is applied on the height and width respectively.Using the correlation analysis of scale and area,the combination of scale transformation is selected to meet the challenge;In the occlusion problem,an extremum difference method is proposed to judge whether occlusion occurs,and then the drift of the template is reduced by adjusting the learning rate.In the problem of finding the local extremum set of the two-dimensional matrix,the original solution method is simplified.And a fast calculation method is proposed;adding a color feature to the feature descriptor increases the feature expression ability.(2)Studying a kind of multi-target tracking technique based on deep learning YOLO detection framework which is combined with Kalman trajectory prediction.Each frame of the image is sent to the detection model,and output the detection result.The Kalman tracking model performs motion modeling on the first two states of one target.When the next frame is tracked,the detection result is used as the observation value,and the Kalman tracking results is used as the predicted value.The optimal matching term is found by the area overlap ratio maximization near the trajectory prediction point,and then the Kalman filter tracker is updated at the same time.As for YOLO detection model,the network was retrained on the PASCAL data set to test the detection performance of the model.(3)Study the optimization of the YOLO detection model.Using the idea of lightweight structure optimization,the traditional volume integral is solved as depth separable convolution and point-by-point convolution,which is applied to the YOLO model,and the two optimization models are compared.Analyze optimize the model to reduce the original model complexity,and do experimental analysis,the final training model weight size reduced from 238 MB to 22.5MB,the model processing speed increased from 19 fps to 30 fps.