Realtime Mask Detection of Kitchen Staff Using YOLOv5 and Edge Computing

We propose a new end-to-end, edge-device embedded solution for detection of improper mask-wearing of kitchen staff using network cameras. Mask detection on kitchen network cameras is a highly specified task with little ready-to use data. The filming position of network cameras and the enviroment in kitchen leaded to vague images and small targets. Also, the detection model should be light-weighted to be deployed on edge computing devices. To improve accuracy on this task, We constructed a novel dataset from real kitchen cameras of different positions, indoor layouts, light conditions and applied effective data augementation. We conducted transfer learning on our dataset starting from COCO pre-trained YOLOV5s weights. In addition, we optimized the model through parameter tuning and post training model pruning for a single Nvidia Jetson Nano device and achieved high accuracy and sensitivity on multiple HD resolution network camera video feed. Experimental results show that our model achieves a training accuracy of 100 percent mAP(0.5) and test accuracy of 97.6 percent mAP(0.5) with minimal training cost: only 89 epochs on our dataset with early stopping. With an inference speed of 10 FPS on Nvidia Jetson Nano, our solution suffices the application requirements and can handle mutiple parallel HD camera streams simultaneously. Compared with previous research, our solution provides competitive cost efficiency where accurate and sensitive high resolution image detection can be run on a single edge device.