Robust Video Multiple Objce Tracking Algorithm Based On Deep Learning Detection

Object detection and tracking are have always been the hot issues in the conputer version.The introduction of convolutional neural networks promotes the development of object detection and target tracking.Multiple object tracking based on detection splits the tracking process into two parts: detection and data association.Most of current detection algorithms have a common shortcoming which is the insufficient use of CNN features.This results in insufficient utilization of smaller size objects information and the detail information contained in the shallow layer.The data association affords many challenges,such as occlusion in the crowded scene and heavy calculation burden caused by the large number of objects.To improve the accuracy of the object detection and data association,therobust video multiple object tracking algorithm based on deep learning detection is proposed.The main research contents are as follows:(1)In order to solve the problem of insufficient use of CNN features,the object detection algorithm based on multi-layer convolution features is proposed.The deep features is upsampled and added with shallow features to produce new features.These new features will be used as the basis in the classifiers.The algorithm achieves good detection results in the VOC2007 data set.(2)In allusion to settle the occlusion in tracking and balance the accuracy and speed of tracking,the multiple object tracking algorithm based on adaptive switching of motion and appearance features.A decision algorithm based on target density is proposed to judge tracking scene.In the complex scene,the appearance feature is used to improve the robustness of the tracking.In the sample scene,the motion feature estimated by RNN is used to accelerate tracking.When the object is occluded for a long time,the estimate of RNN is prone to errors.In order to alleviate the adverse effect,the cascade matching is integrated.(3)In order to verify the effectiveness of the proposed algorithms,the authoritative VOC dataset and MOT dataset is used as evaluation data.The state-of-the-art algorithms is compared with the proposed algorithm in this paper.Experimental data shows that the proposed algorithm has excellent performance and strong competitiveness.