Robot visual perception and autonomous obstacle avoidance based on deep learning

Abstract

Dynamic obstacle avoidance is crucial for a robot to achieve autonomous and safe navigation, especially in complex and ever-changing indoor environments. The robot needs to detect obstacles in a timely manner and dynamically plan a safe path. To accomplish this, an environment perception system is established using an RGB-D depth camera and an IMU unit, providing the robot with multimodal information such as 3D vision and orientation angles. Firstly, an improved target detection model based on YOLOv7-tiny is constructed. This model utilizes the YOLOv7-tiny algorithm to recognize obstacles in color images. The color and depth images are aligned to obtain obstacle size information and spatial distance between the robot and obstacles. The synchronized color and depth information are then inputted into the RTAB-D SLAM algorithm to construct a map. After mapping, the DWA local path planning algorithm is used to plan a path, and obstacle avoidance decisions are sent to the chassis control module to achieve autonomous obstacle avoidance. This enables the robot to navigate autonomously in real-world scenarios. Experimental analysis shows that the improved YOLOv7-tiny target detection algorithm achieves slightly higher accuracy compared to the original algorithm, with a 25% increase in FPS during detection. The proposed method successfully enables the robot to achieve autonomous obstacle avoidance. This research provides a basis and reference for robots to rely solely on visual and inertial sensors for obstacle recognition and autonomous navigation.