A Shallow Approach for Vehicle Speed Estimation in Urban Areas Using YOLO, GOG, and a MLP

Modern traffic is a c omplex c hallenge of a ll urban areas around the globe. The complex nature of traffic m akes it necessary to use specialized tools to understand this phenomenon. Detection, tracking, and speed estimation are computational tools that help understand traffic, providing the necessary insights to design intelligent cities. These three tasks help predict vehicle flow and extract essential information for decision-making and city planning. This paper proposes a system with three components to estimate the speed of vehicles from video of urban scenes. The first c omponent p rocesses t he v ideo f rames u sing t he You Only Look Once (YOLO) algorithm for vehicle detection. The second component uses the Globally-Optimal Greedy Algorithm (GOG) to track the vehicles. Lastly, the third component is a Multi-Layer Perceptron (MLP) that predicts vehicle speed based on differences from frame to frame. Our proposed system was tested by estimating vehicle speed in real and complex scenes to obtain a deeper insight into the behavior of the vehicles. The system is robust enough to work with different car types, light conditions, weather conditions, and camera positions. This paper presents our proposal, shows the experiments, and compares it with similar studies. Also, we draw conclusions and provide directions for future research.