| Autonomous mobile robots are widely used in many fields.Automatic obstacle avoidance is a key problem for robots to achieve fully automatic operations.To ensure the safe driving of robots,it is necessary to have the ability to perceive the environment and avoid existing dangers.Traditional methods are hard to implement correct obstacle avoidance in complex outside scenes,in this paper,a related algorithm of deep learning is used to design a visual obstacle avoidance algorithm based on monocular images,and the effectiveness of the obstacle avoidance algorithm is proved by the omnidirectional robot as a carrier.First,the obstacle targets are detected based on the YOLOv4 target detection algorithm,and road obstacle data sets are collected and produced.The corresponding improvement methods are designed for the YOLOv4 model from the network structure and detection method.Structurally,the commonly used coupling detection head of the YOLO series is abandoned,replaced with a decoupling head to handle classification and regression tasks separately,the confidence is classified as a regression branch,and the convolutional layer is replaced by a fully connected layer to handle the classification problem,and use the convolutional layer to deal with the regression problem.In the detection method,the crossentropy loss function is changed to the Focal Loss function with better performance,and the original NMS is replaced with Soft-NMS to prevent the excessive suppression of the bounding box from causing missed detection.The experimental results show that the improved YOLOv4 model has a better detection effect,which proves the effectiveness of the improved method.Secondly,an obstacle relative distance perception algorithm is designed,and a very effective monocular depth estimation algorithm Monodepth2 is selected to provide scene depth information within the camera’s field of view.Use the obstacle bounding box information obtained by target detection to locate the obstacle coordinate position in the depth map,and apply the pixel value-based relative distance perception(PV-RDP)to obtain the relative distance information of all obstacles in the field of view,which can indicate the positional relationship of all obstacles in the field of view,and provide the necessary information for the obstacle avoidance process.Finally,according to the selected three-wheeled omnidirectional mobile robot’s wheel train structure,the robot’s kinematics model and dynamics model were established respectively.A vision-based obstacle avoidance algorithm is designed,and the obstacle avoidance algorithm is mounted on a three-wheeled omnidirectional robot for obstacle avoidance experiments.Set up multiple obstacles in indoor and outdoor scenes for obstacle avoidance experiments.The experiments show that the visual obstacle avoidance method based on monocular images can perceive the obstacles in the field of vision in real time,not limited by the scene,and can effectively perceive the front.Obstacles and complete obstacle avoidance tasks in real time. |
