Image- and point-cloud-based computer vision that tracks workers across a job site and warns them when they enter falling-object danger zones.

Objective. Fatal injuries on construction sites have risen sharply in recent years. My team devised a method to improve worker awareness and safety on job sites — reducing downtime and preventable accidents.

Solution. We combined image processing and point-cloud vision to track worker movement and notify them when they enter dangerous regions, targeting the most common hazard: being struck by falling objects. First we preprocess the worksite by selecting four fixed reference points in the camera's field of view. Assuming a fixed camera and that workers stand on a common ground plane, we compute a static homography that maps pixel coordinates to world coordinates, accounting for the camera's perspective.

We then segment the video stream into frames for a YOLOv7 network, refined via transfer learning on our construction-site dataset. YOLO detects workers and outputs bounding boxes; we estimate each worker's position from the bottom-center of the box, map it to world coordinates through the homography, and stitch frames into an animation of worker positions over time. Danger areas are predefined, though a network could be trained to flag risky objects automatically.

Results. We render worker locations from a bird's-eye view. Future work includes richer position representations (velocities, behavior prediction), multi-camera coverage, and occlusion handling. Across a variety of environments we successfully tracked each worker's location.