Detecção profunda de semáforo por sobreposição de contexto sintético em imagens naturais arbitrárias

Abstract

The use of deep neural networks as a solution to problems related to autonomous driving has been increasingly considered by the researchers. With this tooling, common traffic elements, such as pedestrians, traffic signs and traffic lights can be detected effectively, by simply providing as input data a representative amount of images that describe a real traffic context. In particular, the detection of traffic lights and the correct classification of their state are essential in preventing accidents. However, collecting and annotating such set of traffic light data can be a highly costly task, both in time and effort. To overcome this problem, it is proposed assembling an expressive dataset that overlaps traffic contexts generated synthetically, through simple computer graphics, on arbitrary images containing natural scenes not related to traffic. This dispenses the need for collection of real-world data, automates the annotation of traffic lights arranged in the generated scene, and also makes it possible to balance the occurrences of the yellow state, which would be difficult to capture, with those of the other states. Experiments revealed that using the method yields results comparable to those obtained using real-world data, with average mAP and F1-score about 4 percent points higher.