Night-time pedestrian classification with histograms of oriented gradients-local binary patterns vectors

The use of night vision systems in vehicles is becoming increasingly common, not just in luxury cars but also in the more cost sensitive sectors. Numerous approaches using infrared sensors have been proposed in the literature to detect and classify pedestrians in low visibility situations. However, the performance of these systems is limited by the capability of the classifier. This paper presents a novel method of classifying pedestrians in far-infrared automotive imagery. Regions of interest are segmented from the infrared frame using seeded region growing. A novel method of filtering the region growing results based on the location and size of the bounding box within the frame is described. This results in a smaller number of regions of interest for classification, leading to a reduced false positive rate. Histograms of oriented gradient features and local binary pattern features are extracted from the regions of interest and concatenated to form a feature for classification. Pedestrians are tracked with a Kalman filter to increase detection rates and system robustness. Detection rates of 98%, and false positive rates of 1% have been achieved on a database of 2000 images and streams of video; this is a 3% improvement on previously reported detection rates.