The improved spatial histogram features and integrated system for vehicle detection

In this paper, we propose an improved spatial histogram features that using object´s geometric information. The templates of spatial histogram features created by superpixel image, which encoding the spatial distributions of objects, instead of the usual random way. In order to promote the precision of vehicle detection, a detection system consists of global-based representation features and part-based representation is employed. The improved histogram features, as global features, feed to a support vector machine (SVM) to make a decision. The candidates area that indicated by global-based representation features are re-detected by the procedure of part-based representation detection. In this procedure, we extract local binary patterns (LBP) descriptor from templates´ windows. SVM boosting method is applied to learn every group of part-based representation features, and then a threshold of accuracy is set to do some group selection work. In experiment on vehicle dataset shows that the improved spatial histogram features is efficient and robust in object detection, and the proposed system, hybrid of global and local information can successfully improve the precision of detection.