Study On Traffic Sign Detection Algorithm Based On Deformable Part-based Models

Traffic sign detection and recognition are important for intelligent traffic system. Traffic sign detection identifies the existence of traffic sign candidates in an image and the recognition stage classifies or rejects the detected candidates. Traffic sign detection is critical because it can give an alarm immediately to the driver and greatly save a lot of time for the following recognition stage. However, traffic sign images are obtained from outdoor natural scenes by camera installed on vehicles, and then, the images are input to a computer for processing. Due to the complicated factors present outdoors, the main difficulties of traffic sign detection are how to extract the traffic signs with various lighting, occlusion, and dimension change and so on. In this thesis, we investigated traffic sign detection and the main works are as follows.1. In order to enhance the traffic signs in the traffic scene images, we designed a method, which integrates both the local color enhancement algorithm base on normalized red-blue channel and global traffic sign enhancement algorithm based on frequency-tuned saliency map. The experimental results show that the method is capable of highlighting the traffic sign, suppressing background and overcoming changes of illumination.2. In the detection stage, traffic signs are divided into five categories, including diamond, triangle, upside-down triangle, circle and octagon. The traffic sign detection framework based on deformable part-base models (DPM) are proposed, which is capable of detecting multiclass traffic sign by their shapes. Single model and mix models of traffic signs are designed, and we get these models by using latent support vector machine algorithm. Experiments shows that the DPM framework is effective in normal traffic sign detection by their shapes.3. The shape, size, area, position and color of occlusion on traffic sign are highly uncertain in traffic scene images. We set up a data set of occluded traffic signs, which contain many occluded traffic sign in different situations. Through experiments with an evaluation data set of high-resolution images, we show that the proposed framework is able to detect occluded traffic signs with high detection accuracy and that it outperforms other approaches in defaced and other degraded conditions.