Lightweight Edge Architecture for Real-Time Object Recognition and Depth Estimation in Autonomous Systems

This paper presents RAPTOR (Real-time Architecture for low Power Targeting and Object Recognition), an innovative energy-efficient, lightweight edge compu-ting architecture that combines object recognition and depth estimation from ste-reo vision for autonomous robots, vehicles, and drones. The computational effi-ciency challenges addressed in this research are particularly relevant given recent industry shifts toward camera-based autonomous systems, as demonstrated by Tesla's transition to vision-only self-driving technology [Gent, 2021]. The pro-posed system leverages parallel feature extraction on dual Sony IMX500 AI cameras integrated with a Raspberry Pi 5 and Google Coral Edge TPU to achieve real-time performance at minimal power consumption. By addressing the compu-tational expense of stereo vision on resource-constrained platforms, our approach proposes a unified feature extraction pipeline that serves both visual perception tasks simultaneously. Preliminary analysis suggests the system can achieve com-parable accuracy to existing methods while consuming significantly less power than current industry solutions, requiring <10 watts compared to Tesla's FSD computer at 72 watts [Bannon et al., 2019] and typical edge devices like the Jet-son Orin NX at 40 watts. This represents a significant advancement in deploy-ment potential for autonomous navigation systems, offering a more sustainable and scalable solution for widespread adoption.