Vision and Laser-Based Mobile Robot Following and Mapping

This paper presents an algorithm called Real-time Vision-based Following Map (RVFM), which enables a mobile robot to perceive and track human body positions in space based on visual perception. Simultaneously, it creates a map of the environment, addressing the autonomous mapping issue for mobile robots in unpredictable environments. The algorithm utilizes the YOLOv5 algorithm for target recognition and optimizes detection results by clustering real bounding box sizes using the DBSCAN algorithm. By optimizing the loss function, detection performance is further improved. For target tracking, a combination of siamese neural networks and RPN network regression is employed, enhancing tracking accuracy and stability. In terms of path planning, the DWA algorithm is introduced, combined with a cost function representing the field of view, to achieve more accurate path planning and keep the target human within the camera's field of view center. For mapping, the algorithm is based on Gmapping and incorporates Kalman filtering and outlier removal techniques to ensure mapping accuracy and reliability. Experimental results demonstrate that the RVFM algorithm provides more accurate human body position information, achieving precise following. It also performs well in indoor path planning and obstacle avoidance, ensuring safety and efficiency during the following process.