Low-Power AI Processing for Smart Surveillance Systems

Abstract

Power efficiency is becoming a design requirement for smart surveillance, as AI-based cameras are more frequently deployed on battery-powered and thermally constrained edge platforms.Power efficiency has become a design requirement for smart surveillance because of the growing deployment of AI-based surveillance cameras on battery-powered and thermally constrained edge platforms. The methodology used in this research was experimental in order to understand the low-power AI processing techniques to enhance smart surveillance application. A prototype AI-based surveillance pipeline was designed with the use of a lightweight computer-vision model on a resource-constrained edge-computing platform, and evaluated the performance of AI processing based on limited computational and energy resources. Optimizations such as quantization, model pruning, light-weight neural network architectures and reduced input resolution were implemented to reduce computational complexity and power consumption, and a traditional AI-processing configuration was compared with optimized low power configurations. To illustrate the reporting procedure implied by this experimental design, hypothetical results rather than hardware-measured results are shown and interpreted as if they were the actual results from on-device experiments. The illustrative pattern was similar to the literature related to low-power edge-AI: when optimization was done in a limited way, the power draw and speed of inferences dropped significantly compared to the base case, and even more when more intense optimizations were performed together.The trend is the same as in the larger stream of work on low-power edge-AI: limited optimizations led to moderate reductions in detection accuracy and significant reductions in power draw and inference speed, while multiple aggressive optimizations together resulted in even greater decreases in detection accuracy. The paper ends by discussing the accuracy–efficiency trade-off for edge-based smart surveillance and future research directions in hardware-validated versions.