An Edge Computing-Based Embedded Traffic Information Processing Approach: Application of Deep Learning in Existing Traffic Systems

Abstract:

To address traffic congestion, this study improved MobileNetv2-YOLOv4 target detection algorithm (MobileNetv2-YOLOv4-K++F) and introduces an embedded traffic information processing solution based on edge computing. We transition models initially designed for large- scale GPUs to edge computing devices, maximizing the strengths of both deep learning and edge computing technologies. This approach integrates embedded devices with the current traffic system, eliminating the need for extensive equipment updates. The solution enables real-time traffic flow monitoring and license plate recognition at the edge, synchronizing instantaneously with the cloud, allowing for intelligent adjustments of traffic signals and accident forewarnings, enhancing road utilization, and facilitating traffic flow optimization. Through on-site testing using the RK3399PRO development board and the MobileNetv2-YOLOv4-K++F object detection algorithm, the upgrade costs of this approach are less than one-tenth of conventional methods. Under favorable weather conditions, the traffic flow detection accuracy reaches as high as 98%, with license plate recognition exceeding 80%.

Key words:Edge computing, Embedded traffic information processing,Deep learning, Model deployment