Autonomous Vehicle Lane Following Network based on Multimodal Sensor Fusion and LiDAR Reconstruction in Extreme Lighting and Snowing Environments

Engineering notes

Key topics: self-driving car, self-driving, autonomous vehicle, motion planning, lidar, point cloud, sensor fusion, radar, planning, 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

The self-driving car industry is gaining attention for its role in motion planning technology. Deep learning approaches have been implemented to plan autonomous vehicles' motion, but their effectiveness depends on the data used in training. Despite advancements, fully autonomous vehicles still face limitations, such as lighting-related issues affecting camera image accuracy, and LiDARs are negatively impacted by challenging weather circumstances like snow in varied snowfall conditions, thus becoming imperative to precisely control the motion of the Autonomous Vehicle. To address this, an architecture based on deep-learning networks is proposed for lane-following on roads, consisting of LiDAR De-snowing, Multi-modal Sensor Fusion, VGG16, and gated recurrent units (GRUs). The Point Reconstruction Network (PR-Net) rebuilds points from neighbors, while the Reconstruction Difficulty Network forecasts reconstruction difficulty. Radar measurements and de-snowed LiDAR point clouds are projected onto the front-view image plane, fed into VGG16 and GRU output layers, and processed into two fully connected layers for steering and speed parameters estimation.

Links and sources

• Official / arXiv page

Comments