Software-Defined Active LiDARs for Autonomous Driving: A Parallel Intelligence-Based Adaptive Model

LiDAR is an indispensable sensor for autonomous driving that can provide precise 3D information about the environment. Among various types of LiDARs, mechanical LiDARs are the most commonly used on vehicles that uniformly perceive the scene utilizing rotating motors. However, accurate perception of foreground objects is considered as the most important task in automotive LiDARs, therefore the current operating mode of mechanical LiDARs wastes a significant amount of sensing resources on the useless background. Besides, the development of LiDAR hardware and software systems is currently split into two independent segments, lacking real-time interaction between physical entities and digital models in cyberspace. To address these issues, we propose software-defined active LiDARs for autonomous driving based on parallel intelligence. Active LiDARs redefine LiDAR's hardware operation through software systems to achieve adaptive sensing resource allocation, constituting a closed loop between physical space and cyberspace. During the working process, it calculates scenario heatmaps with the constructed high-definition maps (HD maps) in cyberspace at first. Then it takes prescriptive control of physical LiDARs based on heatmaps to improve sensing resource utilization. We build two adaptive LiDAR models in CARLA and construct a hardware prototype in the parallel sensing platform, DAWN. Active Point Cloud (APC), a new dataset collected in CARLA, is proposed and a 3D object detection task is selected to demonstrate the effectiveness of active LiDARs. Our experimental results show that active LiDARs can both improve raw data quality and model performance compared with mechanical LiDARs, especially for the perception of distant objects.