Benchmarking of Dual-Step Neural Networks for Detection of Dangerous Weapons on Edge Devices

Nowadays, criminal activities involving hand-held weapons are widespread throughout the world and pose a significant problem for the community. The development of Video Surveillance Systems (VSV) and Artificial Intelligence (AI) approaches have made it possible to implement automatic systems for detecting dangerous weapons even in crowded environments. However, the detection of hand-held weapons - usually very small in size with respect to the Field of View (FoV) of the camera - is still an open challenge. The use of complex hardware systems and deep learning (DL) architectures have mitigated this problem and achieved excellent results, but involve high costs and high performance that hinder the deployment of such systems. In this contest, we present a comprehensive performance comparison in terms of inference time and detection accuracy of two low-cost edge devices: Google Coral Dev board and NVIDIA Jetson Nano. We deployed and run on both boards a dual-step DL framework for hand-held weapons detection exploiting half-precision floating-point (FP16) quantization on Jetson Nano and 8-bit signed integer (INT8) quantization on Coral Dev. Our results show that both in terms of PASCAL VOC mean Average Precision (mAP) and Frames per Second (FPS), the framework running on Jetson Nano (mAP = 99.6, FPS = 4.5, 2.5, 1.7, 1.4 from 1 to 4 people in the camera FoV, respectively) slightly outperform the Coral's one (mAP = 98.8 and FPS = 2.9. 1.5, 1.1, 0.9 from 1 to 4 people in the camera FoV, respectively). The Coral Dev obtained the highest inference speed (FPS = 36.5) overcoming the Jetson Nano (FPS=23.8) only when running the dual-step framework with no people in the camera FoV. In conclusion, the benchmark on the two edge devices points out that both allow to run the framework with satisfactory results, pushing towards the diffusion of such on-the-edge systems in a real-world scenario