Vision-based tracking method of nighttime construction workers by integrating YOLOv5 and Deepsort

Abstract

Due to poor visibility and fatigue factors, although nighttime construction has been widely used, its safety problems like struck-by accidents have also become increasingly prominent. Most of the current tracking methods of workers are not suitable for direct application in nighttime construction scenarios, so this research proposes a vision-based method, which integrates low-light image enhancement technology, YOLOv5 and Deepsort to track nighttime workers. The proposed method is mainly composed of four modules, including illumination enhancement module, detection module, the Kalman filter and matching module. In the experiment based on nine test videos, the method achieved the average multiple-object tracking accuracy (MOTA) of 89.93% and multiple-object tracking precision (MOTP) of 97.07%. At the same time, the experimental results also show that the method is robust to the common tracking challenges of occlusions, scale variations and posture variations. The proposed method has practical application potential in the monitoring task in nighttime construction, which makes the nighttime construction activities safer and more efficient.