Occlusion robust 2D scan matching using template matching and ray tracing

Abstract We propose an occlusion robust 2D scan matching method utilizing template matching and ray tracing. Scan matching is a fundamental technique for mobile robots in many application scenarios, such as localization, loop closure in SLAM, and object detection. Although it is an extensively researched topic, dealing with the occlusion of 2D LiDAR scans remains challenging due to the limited discriminative information in 2D range scans. We therefore leverage ray tracing within SLAM to explicitly label unknown (occluded) vs. empty (free) cells, and incorporate this occlusion information into the matching score. Specifically, we demonstrate the effectiveness of our approach by solving a stop position determination problem, where scan matching is implemented to recognize a reference object in a 2D map and the mobile robot determines its stop position based on the recognized reference object. The reference object template is prepared in advance by cropping the reference object region from a 2D map. When the position of the reference object changes, the robot creates a new 2D map of the updated environment by using ray tracing, detects the template on the new map, and calculates the new stop position based on the detection result. The technical challenge of this strategy is that the shape of the reference object on the map changes when its position or the set of observable sides changes due to occlusion. Therefore, we propose a template-matching-based detection method that takes occlusion into account. By assigning lower penalties to unknown areas while penalizing empty-occupied conflicts, the matcher becomes less sensitive to occlusion and more selective against false positives. In a verification experiment, we confirmed that our algorithm could detect the reference object with occlusion more robustly than a baseline template matching method in a simulated environment. In object recognition with an occlusion rate of approximately 50% to 75%, the proposed method was able to reduce the recognition error rate by 1/5 to 1/3 compared to the baseline method.