A Tracking Method of Multi-Workers Onsite with Kalman Filter and OpenPose

Monitoring onsite construction workers are crucial for safety inspection and project management. However, tracking multi-workers’ continuous dynamic trajectories and behaviors are still full of challenge in computer vision research field. Moreover, unlike automatic unmanned or pedestrian tracking fields, construction onsite tracking lacks labeled surveilling data sets, which leads to low-effective coherent benchmarks for training neural network models. To solve above problems, this paper proposed an associate-method with Kalman filter and OpenPose model to get trajectories, behaviors, and workers’ ID together simultaneously. This method can keep tracking correct ID in complex onsite environment, reduce manual labeling workload, and enhance the preprocessing speed of original onsite videos for tracking datasets production.