Real-time citywide reconstruction of traffic flow from moving cameras on lightweight edge devices

Abstract

Traffic flow estimation is required for road infrastructure management tasks such as road development planning, routing, and navigation. Determining traffic flow on a citywide scale is challenging because of the expensive costs and portability of current devices. Portable sensing devices such as drive cams and smartphones are an effective source of monitoring the road infrastructure environment because of their continuous interaction with the surrounding. However, the use of such devices to estimate real-time traffic flow has not been fully explored. In this study, we optimize the vehicle detection neural network for inference on lightweight edge devices and develop a client–server framework to reduce and share the computational load to make accurate real-time traffic flow processing from moving camera videos. We conduct extensive research work for various input network sizes and frame rates combinations for three widely used edge devices – Jetson Xavier AGX, Jetson Xavier NX, and Jetson Nano. We obtain a traffic flow reconstruction accuracy ranging from 73.1% to 80.8% evaluated using ground truth data. With the proliferation of moving cameras in vehicles (dash cams, stereo cams, etc.) and inexpensive edge devices, we expect our real-time traffic flow estimation algorithm to have a very promising future. Our findings may serve as a useful reference for several domains in the area of real-time artificial intelligence applications and emerging edge computing devices.