Estimation of Traffic Density Using CNN with Simple Architecture

Traffic congestion might be a problem that is commonly encountered in large cities. Currently, most traffic control systems are still unable to capture traffic data, which means that traffic lights cannot be programmed to be adaptive. In this research paper, a traffic density estimation system based on Convolutional Neural Network was created. In order to do so, a video frame from a road surveillance camera was divided into several blocks. The CNN was then used to predict whether each of those blocks was occupied by vehicles. By doing so, the traffic density of each frame is able to be estimated. The result showed that the simplest CNN model, which only consisted of 27,074 weights and biases, achieved the accuracy of 97.47% and 96.57% towards training and validation data, respectively. The processing speed itself is decent since the system was able to run at approximately 15.52 frames per second.