Green Deep Learning: Comparative Study of Road Object Detectors Between Jetson Boards and PC

Recent advancements in deep learning have provided powerful tools for intelligent vehicle tasks, particularly in the field of perception. However, achieving real-time performance with low power consumption remains a challenge included in the hot research topic known as green deep learning. In this paper, we present a comparative analysis of various YOLOv5 weights trained on the KITTI and SHIFT datasets using two platforms with different power consumption profiles: the NVIDIA Jetson AGX Xavier and a desktop computer equipped with a NVIDIA GTX 1080 Ti. Our analysis focuses on the average inference time and precision metrics for road objects detection, a key task for intelligent vehicles. Additionally, we apply TensorRT to optimize and accelerate the architecture on both platforms, resulting in significant speed improvements, particularly on the low-power Jetson AGX Xavier (30W). Our ultimate goal is to implement our whole autonomous driving architecture on several Jetson AGX Xaviers connected to a PC where the hyper-realistic CARLA simulator, is replicating the real-world autonomous vehicle environment. We obtain compelling validation results on KITTI and CARLA, achieving real-time performance on a lightweight Jetson AGX Xavier with a powerful object detector such as YOLOv5m.