Learning Traffic Light Colors

Abstract

Traffic light recognition is of great interest for advanced driver assistance systems and autonomous driving but still an unsolved problem. While a traffic light has few visual features for detection from camera images we believe the characteristic light represents a potentially very strong and stable feature. The traffic light is actively emitting light which is rarely influenced by weather or lighting condition. When using a color lookup table for an image segmentation-based object detector, the process of creating the lookup table is the crucial point. In this paper, we propose a method for generating a lookup table using real world data of a large dataset. The training data is sampled from labeled objects and stored as multisets. We contribute a frequency-based filtering method to clean the samples before using a k-nearest neighbor classifier to generalize. The result is stored as a three dimensional lookup table. The main contribution is a neighborhood-biasing technique that allows setting different operating points online without retraining. A challenging real world dataset containing several thousands of active lights is used to evaluate the process.