Vehicle Speed Monitoring using Convolutional Neural Networks

Recently, Computer Vision Techniques have been pushing the development of robust traffic monitoring systems. Such methods utilize images captured by video cameras to infer important traffic features, such as vehicle speed and traffic density. Frame Subtraction is currently the most used method to detect vehicles in a video stream, but there are scenarios where this method provides poor accuracy, given their struggle in handling disturbances caused by lighting changes, pedestrians in the scene, etc. In order to improve the accuracy of Traffic Monitoring Systems (TMS), this paper proposes a novel TMS design and implementation in which a Convolutional Neural Network is used to replace Frame Subtraction methods in the vehicles detection task. The results show up to 12% improvements on Vehicle Detection in comparison with Frame Subtraction-based systems, proving its effectiveness on challenging scenarios, while maintaining an error rate of 5% for speed detection.