| A set of algorithms are proposed for traffic signal recognition (TSR) on challenging videos. During the method development, minimal to no assumptions were made about the uniformity of cameras, the accessibility of advanced controls (e.g., shutter speed), the availability of camera-dependent sample data, the environmental lighting conditions, or the distance to the traffic lights. Such openness of input requires the algorithms to be relatively generic and adaptable to various devices and scenarios.;The proposed methodology consists of two major subsets: 1) image based traffic light detection and classification and 2) spatiotemporal information based coordination. At the core of the methodology is a candidate traffic light detection method based on the concept of conspicuity, which involves lightness, color saturation, and contrast. Detected candidates are then classified based on robust relative color similarity. When processing a video, spatiotemporal information (i.e., GPS based camera position and frame timestamp) is used to effectively narrow down the temporal search range and coordinate TSR across frames.;Naturalistic driving videos were tested against these algorithms to analyze the performance and reveal challenges. The proposed detection method outperformed two other generic detection algorithms in nearly all lighting-distance scenarios, although the absolute recall rates (around 50%) were low due to the compromised data quality. Classification achieved nearly 95% accuracy even with strong color variation in the data. The spatiotemporal coordination effectively reduced the data and helped to reach ideal temporal accuracy of TSR through persistent tracking. Challenge wise, sunny daytime was found undesirable due to strong ambient light and a single set of parameters in the detection model was not optimal for all lighting conditions. Nevertheless, intuitive rules were found for tuning the model towards different lighting conditions.;In summary, this study contributes to the state of knowledge in TSR by proposing a set of novel algorithms and analyzing their performance on unprecedented naturalistic driving data. These algorithms are expected to be more suitable than existing methods for processing videos acquired by a diverse camera set under various lighting conditions. |
