TinyLid: a RISC-V accelerated Neural Network For LiDAR Contaminant Classification in Autonomous Vehicle

LiDAR plays a critical role in autonomous car perception. Hence, the robustness of LiDAR data is imperative. However, malfunctions resulting from sensor cover contaminants are unavoidable and can lead to erroneous data that slowly degrade performance. As such, detecting contamination in LiDAR is essential but remains an open challenge due to varying contaminant types, changing properties over time, and deployment aspects. Automatic classification of the contaminants would enable the automated response (like cleaning the sensor) to ensure the integrity of the data collected by the LiDAR sensor. To minimize the effect on the whole vehicle perception system, the contamination classification has to be performed near the sensor and in a computationally efficient way. To address these challenges, we have conducted a feasibility study of developing an efficient near-sensor machine learning-powered contaminant classification running on the RISC-V architecture. This paper proposes a lightweight 2D CNN network, TinyLid, trained to classify contaminants based on the most comprehensive LiDAR contaminant dataset. The results presented in this paper show that the proposed solution can achieve high classification performance while being computationally efficient and running on hardware with negligible power consumption compared to the LiDAR sensor itself. Specifically, implementing a proposed ML model on a reference RISC-V architecture GAP8 achieves the inference time of 2.575 milliseconds, 6.138 operations/cycle, and uses only 6.8% of 512 KiB L2 memory. The results presented in this work showcase the possibility of increasing the reliability and integrity of the LiDAR-collected sensor data without significant computational or energy consumption impact on the broader system.