Real-Time Image Detection of Traffic Light in Complex Intersection Traffic Environment

In order to improve traffic safety at the road intersections when vehicles passing by, a traffic light detection technology based on the traditional image processing method for complex intersection traffic environments is proposed. Firstly, a dataset of traffic images was collected, and distribution laws of the HSV pixel values were statistically analyzed to produce a set of segmentation thresholds under some specified probabilities. Secondly, distribution characteristics of the geometric shape parameters for traffic lights in real drive situations were resolved so as to made their numerical intervals certain. Then, with the help of a defined 9-dimensional pixel density feature, a support vector machine was trained to judge whether the image was a traffic light or not and recognize its type, by which way a large number of interference images were filtered out. Finally, another support vector machine trained by a defined 1764-dimensional feature of the extended image regions was used to verify the survived images to reduce the wrong detection rate. Experiment results showed that the detection algorithm worked with a 98.41% accuracy rate and 1.08% missed detection rate, meanwhile showed a strong adaptability and good real-time performance.