Extração de informações em imagens de tráfego: uma abordagem com aprendizado profundo

Abstract

Traffic systems are fundamental in the development of cities. However, these systems increasingly suffer from problems such as congestion. Problems like this can increase fuel consumption and air pollution. In addition, it directly affects people’s health. For example, studies indicate that exposure to traffic is a factor that collaborates with the early stages of myocardial infarction. An efficient way to reduce this problem category is to perform traffic light control intelligently by reinforcement learning or traffic management algorithms. However, it becomes necessary to extract information from the environment to implement this solution category. The advent of digital image processing and convolutional neural networks made it possible to extract data in a less problematic way compared to more traditional methods, such as installing sensors on the tracks. Using images, it is possible to extract different categories of data, such as the number of vehicles in a lane, the time they are stopped, and the category that this work proposes to extract, the origin and destination of vehicles at intersections. With motivation generated from the need to obtain data to solve problems related to traffic, this work contributes with a complete pipeline for image processing in traffic intersections filmed with aerial angle. The pipeline detects vehicles, identifies their trajectories, and metrifies origins and destinations, thus differentiating itself from the researched works in the literature. The pipeline consists of three main blocks. A custom YOLO (You Only Look Once) convolutional neural network capable of detecting vehicles in aerial angled footage. The second block has the implementation of a tracking method referenced in the literature whose objective is to attribute identity to vehicles in all frames. Finally, the third block is called origins and destinations, whose objective is to measure the number of vehicles that pass through a single location in the scene and extract the number of vehicles by the route. As an evaluation method, each block of the pipeline was metric. The detector model reached the result of IDP 77.5% and IDR of 95.8%. The tracking algorithm obtained a result of MOTA 72.6% and MOTP 74.4%. Since each block of the pipeline depends on the other, the overall result is seen through the metrification of the third, "origins and destinations". This step is evaluated in two phases, the first being the number of vehicles that pass through a single point in the scene, where the average OD Error% is 1.80% and the average OD Error% for routes it is 7.53%.