YOLO-CNN – Deep Learning Approach for Vehicle Speed Detection

The tremendous potential benefits of vision-based vehicle speed detection, such as cost reduction and improved extra functionality, are attracting study interest despite all the difficulties and restrictions. According to a recent poll, learning-based methods for solving this issue are still in their infancy. The requirement for a significant amount of data, which must include the input sequences and, more crucially, the output values matching to the real speed of the vehicles, is one of the primary challenges. In order to gather data in this situation, a complicated and expensive system is needed to sync the camera's images with a high-precision speed sensor in order to produce the actual speeds. In this research, we investigate a learning-based method for vehicle speed identification utilizing synthetic images produced from a driving simulator (such as CARLA). We create hundreds of photos with variety according to multiple speeds, various vehicle types and colors, lighting and weather conditions by simulating a virtual camera situated across a stretch of road. CNN-YOLO and 3D-YOLO Tiny, two distinct methods for mapping an image sequence to an output speed (regression), are investigated. We show preliminary findings that demonstrate the approach's significant potential for addressing vehicle speed detection. In fact, the detection accuracy is improved up to 80% which is far better than existing works.