Research Of Traffic Lights Recognition Algorithm Based On Mobile Platform

The traffic scene contains a lot of information, such as road signing, road markings, and traffic lights. Driver's inattention, visual fatigue and misunderstanding will lead to traffic accidents. Especially the information contained in the traffic lights is used to indicate safety and orderly passage of pedestrians and vehicles. As a result, traffic lights'recognition is of great significance for driver assistance or automated driving. Especially with the rapid popularization of smart phones and other consumer mobile terminal, the vehicle assistant driving and automatic driving function integrated with smart phone devices, either for smart phones hardware resources which are increasingly powerful of effective utilization, and it can meet the needs of people's pursuit for safety and convenience. In this paper, we use the applied mathematics knowledge and propose a traffic light recognition algorithm that includes traffic lights'candidate region extraction, identification and temporal information fusion of three stages based on smart phone platform.In the traffic lights'candidate region extraction stage, firstly, an ellipsoid geometry threshold model in HSL color space is built to extract interested color regions. These regions are screened with top-hat transform to obtain candidate regions which is satisfied with color and brightness conditions. Then, in order to save the time overhead of color space transformation, we use the ellipsoid geometry threshold model to judge the color pixels, YUV color (red, green) space pixels' judgment table is established offline. The color space conversion and judgment process are implicit in the judgment table, so the recognition rate of the algorithm is improved.In the recognition process, in order to accurately and quickly identify the status and type of traffic lights, we use dimension-reduced HOG features and Extreme Learning Machine (ELM) to validate these candidate regions. Firstly, two ELM classifiers corresponded to the red lights and green lights are trained offline. Then, HOG features were extracted from these candidate regions, the method of BW dimension reduction and ELM classifier are used to identify the status and type of traffic lights.In the temporal information fusion stage, in order to improve traffic lights status and type recognition's accuracy and stability, firstly, optical flow is employed for tracking and correlation of traffic lights. Then, a multi-frame recognition framework based on a finite state machine is introduced to update the recognition results and make the final decision. Finally, these recognition results are feedback.In this paper, a prototype of the proposed algorithm is implemented on the Samsung 19300 smart phone. Experiment results based on different road environments show that the proposed algorithm can recognize the status and type of traffic lights accurately and rapidly.