Department of Information and Communication Engineering
University of Rajshahi, Bangladesh

Title of the Invited Talk: Person Detection for an Autonomous Construction Vehicle Using Deep Learning and Bayesian Statistical Approaches

Abstract: In autonomous operation of construction vehicle like cranes, detecting objects (such as person) at a construction site using a camera mounted on the crane’s boom is essential. Since the camera moves as the boom shifts position, motion compensation is required to ensure accurate detection. This research proposes a Bayesian statistical framework integrated with the Yolo deep learning model (referred to as Yolo-Bayes) for detecting individuals at construction sites. The motion parameters are estimated using point-to-point feature correspondences and a RANSAC-like algorithm, which selects the best model by randomly sampling points. After compensating for motion, two Bayesian models are used in conjunction with Yolo: one for detecting people in motion when Yolo fails, and another for detecting stationary people missed by Yolo. The motion estimation’s accuracy is validated by generating ground truth images with known translation and rotation changes. Extensive experiments, using a large set of videos from construction sites, demonstrate that the Yolo-Bayes model outperforms the Yolo model.