Proposal of Voxel Semantics Estimation System for Automated Snowplow Operation

A shortage of snow removal workers is a problem in Japan's heavy snowfall regions. In order to solve this problem, there is a need to automate the operation of snowplows. For automation, it is necessary to accurately recognize the surrounding environment, such as snow piles, buildings, and roads, in a three-dimensional space. Therefore, in this study, we propose an accurate environment recognition system using LiDAR and advance map (Semantic Map). The system models the 3D space with Voxels and estimates the semantics of each Voxel by probabilistically integrating the Semantic Map and the Point Cloud Semantic Segmentation, so that the system can be easily integrated with applications such as path planning and snow volume estimation. Appropriate integration of the two types of information can compensate for the recognition errors of each method, resulting in highly accurate environmental recognition. Furthermore, although two semantics estimation methods were integrated in this study, the proposed system can also handle more than three methods, making it an extensible system. Through experiments in snowy environments, we thoroughly analyzed the proposed method. As a result, we found that the building information in the Semantic Map can compensate for the misrecognition of snow piles in the Point Cloud Semantic Segmentation. The results of the quantitative evaluation showed that the proposed method can obtain high IoUs: 73.8% for snow piles, 75.2% for roads, and 83.6% for buildings, which exceeded the accuracy of Semantic Segmentation alone in all classes.