TY  - JOUR
T1  - Implementation of an Edge Computing-Based Industrial Site Monitoring System Using Deep Learning for Object Detection and Tracking
AU  - Kang, Minsung 
AU  - Lim, Youngchul 
JO  - The Journal of Korean Institute of Communications and Information Sciences
PY  - 2025
DA  - 2025/1/1
DO  - 10.7840/kics.2025.50.11.1771
KW  - Edge Computing
KW  - Object Detection
KW  - Multiple Object Tracking
KW  - Industrial Monitoring
KW  - Deep
KW  - Learning Optimization
AB  - This paper proposes a monitoring system for industrial sites by optimizing a YOLOv7-based integrated network for object detection and feature embedding in edge computing environments and applying multi-object tracking techniques. The object detection network was designed with three levels of complexity and three resolutions to meet the diverse requirements of industrial environments. A total of nine scale models were trained on the server and deployed on Jetson Xavier and Nano boards after ONNX conversion and TensorRT-based optimization using FP32, FP16, and INT8 quantization. A multi-object tracking method was developed based on the IOU similarity of detection boxes and the similarity of feature embeddings.
Experimental results showed that the object detection model achieved real-time inference at over 100 FPS on the Xavier board with INT8 operations and up to approximately 70 FPS on the Nano board. In terms of accuracy, INT8 optimization led to an average performance degradation of less than 1%. The multi-object tracking achieved a MOTA of 52.73% with an average inference time of 1.93 ms. Based on models of various scales and object detection and tracking techniques, the proposed system is adaptable to the diverse monitoring needs of industrial environments.