Early Fusion-Based Road Obstacle Risk Classification and Behavior Control Using Camera and 2D LiDAR

Engineering notes

Key topics: autonomous driving, lane detection, lidar, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Implementation of autonomous driving requires location recognition, lane detection, and distance information collection of other vehicles. Information collected from human senses (eyes, ears, touch, etc.) travels to the brain to recognize, judge, and control it. The ultimate goal is to implant this process into the vehicle as it is. Humans recognize distance by calibrating information obtained through both eyes in the brain. To implement this, we collect information about objects using two sensors, camera and LiDAR. Cameras have strengths in object recognition and resolution, and LiDAR has strengths in distance information collection. The fusion of these sensors maximizes their respective strengths, enabling effective detection of obstacles around the vehicle. This is much more effective than systems that rely on a single sensor or whose sensor settings are not properly calibrated or synchronized. Systems based on a single sensor are vulnerable to problems such as object clogging, visually similar items. This can lead to excessive avoidance behavior, unnecessary pauses or obstacle recognition errors. In this work, a combination of camera-LiDAR early fusion, dual-condition gating, and risk-level classification (level 1 = speed bump, level 2 = vehicle, level 3 = human) can improve avoidance success rate and driving consistency compared to single sensor-based control algorithms while reducing unnecessary responses.

Links and sources

• Official / arXiv page

Comments