Object detection in video surveillance using MobileNetV2 on resource-constrained low-power edge devices

Abstract

Edge-based video surveillance systems encounter significant obstacles in object detection due to the limited computational power and energy efficiency of edge devices, which are required to deliver real-time processing capabilities. Traditional object detection models are excessively resource-hungry for these environments, making optimization techniques absolutely essential. This study robustly explores the implementation of quantized transfer learning utilizing SSD MobileNet V2 with 8-bit quantization to significantly elevate the performance of object detection on resource-constrained edge devices. Experimental results decisively indicate that the Raspberry Pi 5 and Nvidia Jetson Orin Nano surpass other devices, achieving total latencies of 5 ms and 85 ms, respectively, affirming their exceptional suitability for real-time applications. The quantized int8 model secures an impressive accuracy of 80.65% while dramatically lowering memory consumption and latency when compared to the unoptimized int32 model. Furthermore, the model demonstrates outstanding performance on a masked-unmasked dataset with an F1 score of 0.92 for masked detection. These findings underscore the transformative potential of quantization in enhancing both inference speed and resource efficiency in edge-based surveillance systems. Future research will boldly investigate advanced hybrid quantization strategies and architectural enhancements to achieve an optimal balance of efficiency and accuracy, alongside scalability across a broader spectrum of edge devices and datasets.