Traffic light recognition with HUV-histogram from daytime driving-view images

Signals of traffic lights are important factors in traffic environment analysis. Advanced Driver-Assistance Systems (ADASs), which need to be aware of surrounding traffic environments to assist drivers, are required to be able to recognize signals of traffic lights. We propose robust, Korean-standard-based, computerized methods to recognize traffic lights in daytime driving-view images. The images captured by the camera mounted on the vehicle. To detect traffic lights in the images, we use the Aggregate Channel Features (ACF). To recognize states of traffic lights, we remove irrelevant background pixels in the detected traffic light image patches by using visual saliency analysis, then we train support vector machine (SVM) with HUV-histogram. Our methods do not require any specific sensors such as Lidar or external systems such as car-to-car networks. This makes our methods to be easily ported over to existing automotive control platforms. 13,500 frames of videos were used to test our methods. The traffic light recognition accuracy is 0.97.