Object Detection Algorithm of Stairs Based on Multi-information Fusion

In view of the mobile robot's perception of the environment in the process of moving, the method of deep learning is used to detect the target terrain in order to send follow-up instructions to the robot. In this paper, a lightweight stairs detection algorithm based on improved YOLOv5-Lite is proposed, which uses RepVGG as the backbone network to extract features to build YOLOv5-Lite lightweight neural network model, and removes the Focus layer in the backbone network and replaces it with a 6x6 convolution layer to reduce the number of network parameters to improve the speed of model detection. Through the introduction of BasicRFB_s pooling layer, increase the receptive field of the network, and add C3SE attention mechanism in the pooling layer to reduce pooling loss. To decrease the mistakes, we also segment the point cloud of stairs into vertical and horizontal planes by the difference of normal and consider their probability distribution for the further detection. The experimental results show that the accuracy of the improved lightweight algorithm can reach 97.1%, and the frame rate can reach 51 fps. After segmenting the point cloud, the accuracy of detection method can reach 99.9% and the frame rate can reach 46 fps. The results show that the proposed method has the characteristics of real-time, lightweight and high accuracy.