Toward Neuromorphic Perception: Spike-driven Lane Segmentation for Autonomous Driving using LiDAR Sensor

Abstract

As a prerequisite of high vehicle autonomy, lane segmentation is a significant perception task for advanced autonomous driving. In recent years, spiking neural networks (SNNs) have garnered the attention of researchers due to their appealing power efficiency, which provides the potential to improve energy consumption for the perception system on power-constrained autonomous vehicles. In this paper, we propose a spiking neural network targeted for LiDAR sensors to solve the lane segmentation problem. By encoding the LiDAR point cloud into spikes, the proposed SNN constructed in an end-to-end fully convolutional network structure is capable of processing the LiDAR input through the network to segment the lane area effectively. Experiments conducted on the KITTI dataset for urban scenes and the power consumption evaluation demonstrate the high performance and energy efficiency of the proposed SNN for LiDAR-based lane segmentation.