A Robust Pedestrian Detector Based On Heterogeneous Features Fusion

The research of pedestrian detection is one of the hotspots in the field of computer vision It has a broad application prospect in the field of intelligent monitoring systems and driver assistance systems. In recent years, because of the highly frequently occurred traffic accidents how to avoid or reduce traffic accidents has become the focus of public concern. Pedestrians are non-rigid objects and would appear in the scenes with various styles of clothing and complex backgrounds. There also exist many objects with similar appearance to pedestrians in the road area. All the factors above have increased the difficulty and complexity of the pedestrian detection.In this paper, monocular vision sensor is selected as the primary means of environmental information acquisition, and a pedestrian detection system is created to be applicable under the complex conditions. Pedestrian detection has important applications in driver assistance systems. The detection performance is impacted by weather, clothing, lighting, occlusion and other factors. Aiming at solving these challenges, in this paper, first, a new color moments feature is presented to describe the local similarity structure of pedestrians, which reduces the influence of complicated background. Then, a combination coefficient method is introduced to effectively fuse three heterogeneous features, COLOR, HOG, and LBP, which makes better use of each feature. Finally, pedestrians of various poses and views are divided into subclasses with S-Isomap and K-means algorithm. A classifier is trained for each subclass. With respect to the output values of different subclass classifiers, an equally weighted sum based multi-pose-view ensemble detector is proposed. Experiment results on public datasets demonstrate that the proposed feature combination method significantly improves the description capabilities of pedestrian features. Compared with the existing methods, the proposed detector combining the feature and multi-pose-view ensemble detector boosts the detection accuracy effectively.