Accelerating point cloud analytics on resource-constrained edge devices

3D object detection is crucial in various applications, particularly in the fields of autonomous driving and robotics. These applications are typically installed on edge devices to quickly interact with the environment and often necessitate nearly instantaneous reaction. Executing 3D detection on the edge using complicated neural networks is daunting due to the constrained computational resources. Conventional methods like offloading tasks to the cloud result in substantial delays because of the extensive volume of point cloud data being transmitted. In order to address the conflict between constrained edge devices and demanding inference tasks, we investigate the potential of empowering rapid 2D detection to extrapolate 3D bounding boxes. To achieve this goal, we introduce Moby, an innovative system that showcases the practicality and promise of our methodology. We propose a lightweight transformation to efficiently and accurately produces 3D bounding boxes using 2D detection results, eliminating the need for heavy 3D detectors. In addition, we develop a frame offloading scheduler that determines the optimal timing to activate the 3D detector in the cloud, preventing the accumulation of errors. Our evaluations conducted on the NVIDIA Jetson TX2 using real autonomous driving dataset show that Moby provides a latency improvement of up to 91.9% with only minimal decrease in accuracy.