Deep Learning for Risk Detection and Trajectory Tracking at Construction Sites

This paper investigates deep learning for risk detection and trajectory tracking at construction sites. Typically, safety officers are responsible for inspecting and verifying site safety due to many potential risks. Traditional target detection algorithms depend heavily on hand-crafted features. However, these features are difficult to design, and detection accuracy is poor. To solve these problems, this paper proposes a deep-learning-based detection algorithm that uses pedestrian wearable devices (e.g., helmets and colored vests) to identify pedestrians. We train a special dataset by labeling helmets and colored vests to detect the two features among construction workers. Specifically, Kalman filter and Hungarian matching algorithms are employed to track pedestrian trajectories. The testing experiment is run on an NVIDIA GeForce GTX 1080Ti with a detection speed of 18 frames/s. The mean average precision can reach 0.89 when the intersection over union is set at 0.5.