CNN-Planner: A neural path planner based on sensor fusion in the bird's eye view representation space for mapless autonomous driving

• Luis Rosero USP
• Júnior Silva USP
• Denis Wolf USP
• Fernando Osório USP

Resumo

Autonomous driving in urban environments without High Definition(HD) maps is probably one of the most difficult tasks for Autonomous Vehicles (AVs), Many solutions grouped in modular pipelines and end-to-end driving approaches have been proposed for solving this motion planning problem. However, safe mapless autonomous driving has not been fully resolved. In this paper we take the best of these two approaches and propose a motion planner powered by deep learning, sensor fusion, and a modular perception pipeline to perform confident mapless autonomous driving. For this, we divide motion planning into two tasks: lateral planning and longitudinal planning. We address lateral planning as a path-following problem and solve it by creating an end-to-end neural path planner that regresses a dense set of waypoints taking as input sensor data fusion in Bird's Eye View (BEV) space, then a low-level MPC controller follows generated waypoints. On the other hand, for longitudinal planning we use a modular pipeline that uses a Finite State Machine (FSM) to make decisions about the ego car speed, where a detection stack defines the FSM inputs and a low-level PID controller follows the desired speed. Putting lateral and longitudinal planners together we obtain a successful motion planner that outperforms several baselines and matches results of other more complex state-of-the-art motion planners in the CARLA Autonomous Driving Leaderboard.

Palavras-chave: Navigation, Pipelines, Urban areas, Data integration, Sensor fusion, Robot sensing systems, Planning