Multi-Input All-Weather Streetlight to Reduce Carbon Footprint of Illuminating Road Infrastructure

Street lighting systems account for 15-40% of the total energy spent in cities worldwide. An ideal energy efficient street lighting system should ensure driver safety during the night with good visibility while reducing energy uses and thus costs. This project proposes a solution using a vision-based vehicle detection to respond to real-time road and traffic conditions. The system comprises of a vehicular detection model comprised of a computer vision object detection model that can be integrated into existing surveillance systems, such as a closed-circuit television (CCTV) camera, which allows for the automatic control and dimming of LED streetlights. The computer vision object detection model uses the YOLOv8 architecture, allowing for massively decreased computational requirements and model complexity while allowing for improved model accuracy. The developed model outperforms other models similarly trained in vehicular detection, such as a model from a 2023 paper based on YOLOX. The developed model has an 89% decrease in computational requirements, requiring 14.3 GFLOPS compared to 131.0 GFLOPS, and it has a 79.85% decrease in model complexity, with 11.14 MParams compared to 55.31 MParams. The developed model also performs better with a 4.12% increase in accuracy, with a mAP of 0.961 compared to 0.923. Decreased computation requirements and model parameter count allows the model to run on more affordable hardware, allowing a more widespread adoption in communities. More installations of the developed solution means that the energy consumption of streetlights can be rapidly decreased, thus providing long lasting environmental benefits.