Resource-Constrained Onboard Inference of 3D Object Detection and Localisation in Point Clouds Targeting Self-Driving Applications

Research about deep learning applied in object detection tasks in LiDAR data has been massively widespread in recent years, achieving notable developments, namely in improving precision and inference speed performances. These improvements have been facilitated by powerful GPU servers, taking advantage of their capacity to train the networks in reasonable periods and their parallel architecture that allows for high performance and real-time inference. However, these features are limited in autonomous driving due to space, power capacity, and inference time constraints, and onboard devices are not as powerful as their counterparts used for training. This paper investigates the use of a deep learning-based method in edge devices for onboard real-time inference that is power-effective and low in terms of space-constrained demand. A methodology is proposed for deploying high-end GPU-specific models in edge devices for onboard inference, consisting of a two-folder flow: study model hyperparameters' implications in meeting application requirements; and compression of the network for meeting the board resource limitations. A hybrid FPGA-CPU board is proposed as an effective onboard inference solution by comparing its performance in the KITTI dataset with computer performances. The achieved accuracy is comparable to the PC-based deep learning method with a plus that it is more effective for real-time inference, power limited and space-constrained purposes.