Automatic Traffic Light to Ego Vehicle Lane Association at Complex Intersections

We introduce an approach that enables associating all relevant traffic lights to the ego vehicle lane at complex intersection scenarios. To achieve this, we use a combination of image data and traffic light metadata. First, we prepare the input image data that consists of a partly Inverse Perspective Mapping (IPM) for the road and highlight traffic light metadata information like position and colour in the image. Next, we train with the input data a common Convolutional Neural Network (CNN) topology on a target regression vector. This vector contains the column positions of all relevant traffic lights in the image. Finally, we perform two post-processing steps that improve single traffic light results. Results on average yield an accuracy of 86.40%. In the near field, accuracy improves to 93.88%. Moreover, we carry out a subjective test, where humans are presented with the same images as processed by the CNN. This defines a reliable baseline for detailed comparisons against our results. In sum, our approach and the subjective test achieve similar results on average. In detail, the subjective test reaches more accuracy for highly complex intersection scenarios. However, it has less accuracy for distances between 25 m to 50 m to the stop line than the proposed approach.