Real-Time Traffic Response Processing Using Convolutional Neural Network and Edge Computing

Abstract

The present state of art is static traffic signals in most cities, whereas some cities use approach based on dynamic traffic signaling using edge detection, vehicular count using object detection, and traffic density estimation which are less efficient. We propose an edge intelligence-based solution which is more efficient in dynamically adjusting the traffic signals. We begin by collecting data and presenting an approach to label them automatically using transfer learning and K-means clustering, followed by manual corrections. We then build and train a custom ResNet model which takes traffic images as input from cameras and determines the traffic density from the same, based on which the green signal time is adjusted in real-time using predetermined parameters. To build our edge computing system, we use Node-RED which is a low-code, flow-based tool that significantly speeds-up development time. This allows us to perform all the above computations in an edge device (say, Raspberry Pi) for low latency. It also removes the need for persistent Internet connection unlike cloud computing, thus enabling real-time processing. The test results for our ResNet model show an accuracy of 84% which is comparable to other existing solutions while being computationally much cheaper.