Multi-Modal Neural Feature Fusion for Automatic Driving Through Perception-Aware Path Planning

Path planning is a significant and challenging task in the domain of automatic driving. Many applications, such as autonomous driving, robotic navigation and aircraft object tracking in complex and changing urban road scenes, need accurate and robust path planning by detecting obstacles in the forward direction. The traditional methods only rely on the path search method without considering the environmental factors, the vehicle path planning method cannot deal with the complex and changeable environment. To deal with above problems, we propose a perception-aware based multi-modal feature fusion approach that combines visual-inertial odometer (VIO) poses and semantic obstacles in the forward scene of vehicles to plan driving paths. The proposed method takes environment awareness as the guide and combines path search algorithm to realize path optimization task in complex environment. The proposed approach first uses a long short memory network (LSTM) to build a VIO that fuses visual and inertial data for pose estimation. To detect obstacles, the proposed method uses a segmentation model with a lightweight structure to extract semantic 3D landmarks. Finally, a path search strategy combining an A* algorithm and visual information is proposed to plan driving paths for intelligent vehicles. We estimate the proposed path planning method on assimilated scenes and public datasets (KITTI and Cityscapes) by using a micro controller (Jetson Xavier NX) installed on a small vehicle. We also show comparable results with path planning that only uses the greedy algorithm or heuristic algorithm without using visual information and show that our method is adequate in coping with different complex scenes.