Weak Appearance Multi-object Tracking Over Cross-domain And Its Embedded Implementation

Multi-object tracking with weak appearance based on the surveillance video is an important issue for intelligent biological laboratory.In these laboratories,the staffs must wear uniform.In the videos captured by the cameras,the similarity of the object appearance features is high,showing a weak appearance characteristic.Because of some factors,such as occlusion,missing detection or false detection are prone to cause tracking failure.In addition,computational cost of deep learning is too high to be realized on embedded platforms.Therefore,we attempt to deploy the multi-object tracking algorithm over cross-domain based on deep learning to the Jetson nano.We aim to achieve multi-object tracking over cross-domain under weak appearance features for special biological laboratory,and deploy this algorithm to embedded devices.In order to improve the reference time of the algorithm under the limited hardware resources,a new lightweight object detection network is generated by model pruning;in order to improve the accuracy of the object tracker,the convolution neural network is used to extract features to achieve multi-object tracking,and the detection results are corrected based on the tracking results combined with the spatiotemporal context information;in order to achieve cross scene target tracking,we combines the temporal information between images to make logical judgment.Finally,the algorithm is deployed on the Jetson nano,and TensorRT is used to further reduce the occupancy of computing resources and improve the reasoning speed of the algorithm.Our main works and innovation are as follows:(1)We proposed a new lightweight object detection network based on YOLOv3.Combined with model pruning,we propose a pruning algorithm based on normalized layer weight evaluation to prune the channels and layers of the network,compress the network parameters,and improve the reference time of the network on the premise of ensuring the detection accuracy.(2)We proposed a multi-object tracking and object corrected algorithm.The algorithm combines convolution neural network to extract feature vectors,is realized by IOU(intersection over union)and cosine distance operation.Besides,we proposed an algorithm that the detection results are corrected based on the tracking results to improve the performance of the tracker.We mainly use the IOU between the two frames before and after the detection of the object for data association.When a new object appears,the feature vector extracted by convolution neural network is used to measure the similarity with the feature vector set,and the corresponding ID is assigned.At the same time,this algorithm uses the existing tracking results,combined with spatiotemporal context information,to correct the current image detection results,improve the accuracy of detection,indirectly improve the accuracy of the tracker.(3)We designed a multi-object tracking system across scenes.We deploy the algorithm to many embedded devices,use TensorRT to improve the reference time of the network.Then the multi-object tracking results obtained from each device are transmitted to the server at the same time.Using the time sequence information between images,cross scene multi-target tracking based on embedded platform is realized under weak appearance features.The experimental results show that the lightweight multi-object tracking algorithm we proposed achieves better performance compared with other multi-object tracking algorithms.What’s more,it maintains higher network compression ratio with less precision loss,which is suitable for the front-end implementation of embedded platform.