Point Cloud-Based Intelligent Collision Avoidance for Stacker Cranes in Stereoscopic Warehouses

Abstract

In stereoscopic warehouses, stacker cranes might collide with obstacles, resulting in substantial economic losses. Therefore, autonomous collision avoidance is essential for the safe operation of stacker cranes. Point cloud-based three-dimensional(3D) object detection networks can provide precise environmental information for devices, among which voxel-based networks are more suitable for scenarios requiring real-time response. However, voxel-based networks perform poorly in detecting small obstacles and situations where multiple obstacles are located at the same bird's-eye view(BEV) coordinate due to voxelization and feature compression along the height. In this work, we propose a point cloud-based intelligent collision avoidance system. We first design and collect the Stereoscopic Warehouse Obstacle Dataset(SWOD), a point cloud dataset containing obstacles. Then, we propose a pillar-based detection network that includes two innovative improvements: density-aware feature augmentation(DAFA) and the Z-axis Repositioning Module(ZRM). Kernel density estimation is used to obtain the probability density of points within each pillar for feature augmentation to enhance the ability to extract features of small obstacles. The ZRM readjusts the z-axis positions of proposals during post-processing to identify multiple obstacles at the same BEV coordinate. Lastly, if the perception results indicate a risk of collision, an emergency braking signal will be sent to the stacker crane. On the SWOD validation set, our proposed network achieves state-of-the-art detection accuracy while maintaining real-time detection speed. Additionally, we verify the feasibility of the collision avoidance system in a real-world scenario.

Keywords: Point cloud-based three-dimensional object detection, Collision avoidance, Intelligent stereoscopic warehouses