Research On Visual Multiple Object Tracking Algorithm Based On Joint Detection

Object detection and multiple object tracking are two very important topics in the field of computer vision.They have important applications in intelligent security,urban transportation,and national defense.With the vigorous development of deep learning,multiple object tracking algorithms based on object detection have also developed rapidly.At present,mainstream multiple object tracking algorithms often treat detection and tracking as two independent models,using a convolutional neural network for feature extraction of target detection tasks,and then using an additional convolutional neural network to extract the target's identity embedding features.Although this approach can ensure that both links have high accuracy,it produces a large amount of repeated calculations and causes the loss of inference speed.At the same time,the improvement of the target detection algorithm often improves the tracking effect.Therefore,the main research contents of this article are as follows:(1)Use a target detection algorithm based on the center point of the target.The current mainstream target detection algorithm is based on the anchor point target detection algorithm.Although this method has good detection results,it is not suitable for cooperating with multiple object tracking tasks because of the effectiveness of the target identity embedded features extracted at the anchor point.It is much smaller than the target identity embedded feature extracted at the target center.At the same time,different anchor points may be responsible for estimating the same target,which will also cause network ambiguity.Therefore,this paper proposes a target detection algorithm based on the target center point,which uses the heat map response of Gaussian kernel function mapping to determine the target center position to avoid the problems caused by anchor points.(2)Aiming at the problem of separation between the detection and tracking of the appearance feature extraction of the current multiple object tracking algorithm,an integrated network design method is proposed.The feature extraction part of the detection and tracking is combined,and a backbone network is used to extract the depth features of the picture.In the way of multi-task learning,the extracted features are given to the two parallel branches of target detection and identity embedding feature extraction,which reduces the repetitive calculation of the convolutional neural network to extract the target features and improves the inference speed.(3)Design the backbone convolutional neural network with residual neural network as the main body.On the basis of Res Net34 network,introduce the idea of deep aggregation network,introduce jump connections between high and low level feature maps for feature fusion,and introduce deformable Convolution makes the size of the receptive field have the ability to adjust adaptively.On the basis of this network,the algorithm variants of the Res Net34 network and HRNet network as the backbone convolutional neural network are designed to adapt to different tracking task scenarios.(4)A data association algorithm that combines appearance features and motion features.The features extracted by the identity embedding feature extraction network are used to describe the appearance characteristics of the target,the Kalman filter algorithm is used to describe the movement information of the target,the two feature description methods are merged to perform tracking data association,and the Hungary algorithm is selected to detect targets in different frames.The matching completes the tracking task.