Ghost Detection Using Spatial Features-Intensity Fusion for Autonomous Mining Trucks

The multi-dust environment in open-pit mines presents a great challenge for the perception of autonomous mining trucks, leading to a large number of ghost detections (i.e., dust is detected as an obstacle), which seriously affects the production efficiency and increases the risk of accidents involving other traffic participants. To solve the ghost detection problem, this paper proposes a Spatial Features-Intensity Fusion method based on LiDAR point clouds collected from real mining areas. This method first constructs an offline data bank of obstacles and dust, in which the prior feature distributions of surface distance, angle, area, and reflected intensity of point clouds are analyzed. In the online classification pipeline, the grid clustering algorithm extracts point cloud clusters in a specific experimental scenario. Then the similarity between features of extracted point cloud clusters and prior feature distributions in the data bank is calculated using the Mahalanobis distance, which determines whether the current point cloud cluster is a ghost detection. The results show that for all detected point cloud clusters of the same category, the classification accuracy of the proposed method achieves 100% for obstacles and 96% for dust. Moreover, misjudging little dust as an obstacle will not lead to risks, and in most cases does not significantly affect the transportation efficiency of autonomous mining trucks. This method has the potential to help autonomous mining trucks eliminate interferences from adverse weather conditions such as dust, rain, snow, and fog, which is necessary for driving safety and efficient production in autonomous mining areas.