Multi-target Tracking with EmbedMask and LSTM Model Fusion

Abstract

The problem of occlusion occurs during multi-target tracking may result in loss of characteristics of tracking target and thus lose the tracking targets. This paper proposes a multi-target vehicle tracking algorithm based on fusion of Embedding Coupling for One-stage Instance Segmentation (EmbedMask) and Long Short-Term Memory (LSTM) model. Firstly, the obtained real-time video data is input into EmbedMask target detection model by frame for target detection. The targets are separated from background, and traditional rectangular box detection is replaced by instance segmentation. Secondly, the maximum feature data of targets is generated by the resent convolution network, which is input into the LSTM model. The continuous data of targets is obtained by calculating and estimating the motion attitude of the tracking target. Finally, the motion and detection data of targets is input into new LSTM model layer, and the fusion calculation is used to reduce the tracking loss caused by overlapping, which can ensure the accuracy of target tracking. Experimental results on standard MOT data sets show that the proposed algorithm is robust and can be used to accurately track occluded overlapping targets.