Link-Aware Frame Selection for Efficient License Plate Recognition in Dynamic Edge Networks

Abstract

With the booming development of Internet of Things (IoT) and computer vision technology, running vision-based applications on IoT devices becomes an overwhelming tide. In vision-based applications, the Automatic License Plate Recognition (ALPR) is one of the fundamental services for smart-city applications such as traffic control, auto-drive and safety monitoring. However, existing works about ALPR usually assume that IoT devices have sufficient power to transmit the whole captured stream to edge servers via stable network links. Considering the limited resources of IoT devices and high-dynamic wireless links, this assumption is not suitable for realizing an efficient ALPR service on low-power IoT devices in real wireless edge networks. In this paper, we propose a link-aware frame selection scheme for ALPR service in dynamic edge networks aiming to reduce the transmission energy consumption of IoT devices. Specifically, we tend to select a few key frames instead of the whole stream and transmit them under good links. We propose a two-layer recognition frame selection algorithm to optimize the frame selection by exploiting both the video content variation and real-time link quality. The extensive results show that, by carefully selecting the offloaded frames to edge servers, our algorithm can significantly reduce the energy consumption of devices by 46.71% and achieve 97.95% recognition accuracy in the high-dynamic wireless link of the edge network.