Coarse-to-fine Method for Vision-based Pedestrian Traffic Light Detection

Abstract

Pedestrian traffic light detection is an important technique of the navigation system for the visually impaired during road crossing. In this paper, a three-stage coarse-to-fine method for pedestrian traffic light detection is proposed. The proposed method is mainly divided into two processes, the training process and the detection process. In the training process, the Gaussian mixture model (GMM) is adopted to determine the parameters of the filter on stage I. The classifier on stage II is trained by a modified convolutional neural network (CNN) to capture features in each channel of the CIELAB color space. The classifier on stage III is trained by the adaptive boosting (AdaBoost) algorithm with Haar features. In the detection process, firstly the board filter is adopted to generate candidate regions of pedestrian traffic lights. Secondly, these candidate regions are detected in multiple scales by the CNN-based classifier with fixed size. Finally the AdaBoost-based classifier is adopted for refinement detection. Testing results verify the effectiveness of the proposed method.

Keywords

• gaussian mixture model
• multi-layer neural network
• boosting
• object detection
• computer vision