RAD: Realtime and Accurate 3D Object Detection on Embedded Systems

Hamed H. Aghdam, Elnaz J. Heravi, Selameab S. Demilew, Robert Laganiere; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 2875-2883

To our knowledge, the fastest 3D object detector on LiDAR data works at 42.03 point clouds-per-second on high-end machines and 6.15 point clouds-per-second on embedded boards. We propose a deep 3D object detector with higher detection accuracy running at 84.46 point clouds-per-second on high-end machines and 10.91 point clouds-per-second on computing boards that is 2 and 1.77 times faster compared to fastest published network. We achieve considerably higher processing rate without reducing the complexity of the network but by designing a more efficient decoder. Our extensive and practical experiments reveal that the detection accuracy of our proposed network is comparable to the best-performed method using practical metrics but it is 3.36 times faster. Besides, we carefully analyze the model and indicate that negligible error in two regression outputs contributes to the reduction in the average precision. Overall, considering the accuracy and speed, our proposed network is highly practical to be executed on embedded boards for ADAS applications.