关键词: 目标检测, 轻量级神经网络, 多目标跟踪, 长短时记忆（LSTM）, YOLOv3

Abstract:

Considering the real-time and identity hopping problems in the process of multi-object tracking, a multi-vehicle tracking algorithm based on detection is proposed. Using Mobilenetv2 to replace the backbone network of YOLOv3 detection algorithm, the object detection module Yolov3-Mobilenetv2 is constructed to reduce the detection algorithm model parameters and improve the running speed of the detection module. Bottom-up connection is introduced in Mobilenetv2 to enhance information fusion among multi-scale feature maps. The LSTM motion model is adopted to solve the prediction error generated by Kalman filtering in the nonlinear system. The LSTM motion model is introduced based on the Deepsort tracking algorithm to form the L-Deepsort tracking algorithm. It improves appearance matching strategy of L-Deepsort tracking algorithm to enhance the correlation between targets. The lightweight object detection algorithm Yolov3-Mobilenetv2 is fused with the multi-object tracking algorithm L-Deepsort to form the MYL-Deepsort multi-vehicle tracking algorithm and realize the real-time and accurate tracking of multi-vehicles. Experimental results show that the speed of this method is 21 frame/s higher than YOLOv3-Deepsort and 13 frame/s higher than YOLOv3-Deepsort on TX2 platform when tracking performance is improved.

Key words: object detection, lightweight neural network, multi-object tracking, Long Short-Term Memory（LSTM）, YOLOv3