Ultra-Sonic Sensor based Object Detection for Autonomous Vehicles

Abstract

Perception systems in autonomous vehicles (AV) have made significant advancements in recent years. Such systems leverage different sensing modalities such as cameras, LiDARs and Radars, and are powered by state-of-the-art deep learning algorithms. Ultrasonic sensors (USS) are a low-cost, durable and robust sensing technology that is particularly suitable for near-range detection in harsh weather conditions, but have received very limited attention in the perception literature. In this work, we present a novel USS-based object detection system that can enable accurate detection of objects in low-speed scenarios. The proposed pipeline involves four steps. First, the input USS data is transformed into a novel voxelized 3D point cloud leveraging the physics of USS. Next, multi-channels Bird Eye's View (BEV) images are generated via projection operators. Later, the resolution of BEV images is enhanced by means of a rolling-window, vehicle movement-aware temporal aggregation process. Finally, the image-like data representation is used to train a deep neural network to detect and localize objects in the 2D plane. We present extensive experiments showing that the proposed framework achieves satisfactory performance across both classic and custom object detection metrics, thus bridging the usecase and literature visibility gap between USS and more established sensors.