Traffic Sign Detection with Occlusion Handling

Abstract

In advanced driver assistance systems (ADAS), accurate detection of traffic signs plays an important role in extracting information about the road ahead. However, traffic signs are persistently occluded by vehicles, trees, and other structures on road. Performance of a detector decreases drastically when occlusions are encountered especially when it is trained using full object templates. Therefore, we propose a new method called Discriminative Patches (d-patches), which is a traffic sign detection framework with occlusion handling capability. Discriminative patches are those regions of an object that possess the most discriminative features compared to their surroundings. These patches are mined during training and are used for classification instead of the full object templates. Secondly, based on our observations, red and blue color in Lab color space can be found at high variance with respect to the mean. By exploiting this we propose a coarser-to-fine approach, which speed-up the over-all traffic sign detection system by four folds without lose in accuracy. We also found that the distribution of redundant-detections around a true positive is different from that around a false positive. We also propose a novel hypothesis generation scheme that uses a voting and penalization mechanism to accurately select a true positive candidate on the basis of its confidence-score and distribution properties, instead of solely relying on the confidence-score value. The proposed method achieves 100% detection accuracy on all three superclasses (i.e., Prohibitive, Mandatory and Danger) of the German Traffic Sign Detection Benchmark. We also introduce a new Korean Traffic Sign Detection (KTSD) dataset, with several evaluation settings to facilitate the performance evaluation of a detector under different conditions. Our proposed algorithm on-average achieves 4.0% better detection accuracy, when compared with other well-known methods, on KTSD dataset, in all three superclasses under partially occluded settings. On Swedish Traffic Sign Detection dataset our method manages to achieve state-of-the-art performance.