Pedestrian detection using structure-constrained features with ada-boost algorithm

Abstract

Among various object detection tasks, pedestrian detection task is especially important since it is highly related to many industrial applications such as autonomous car and surveillance. A pedestrian detector should be able to detect the full-body rectangular regions of humans in a 2D RGB image. It is very challenging task to detect pedestrians in various poses, scales, and lightning conditions. The proposed detector uses features which are semi-automatic. Their structures are defined by finely-designed feature kernels. The shapes and sizes of the kernels are determined with some constraints which reflect various invariance characteristics. However, their positions are discriminatively learned through the Ada-boost algorithm. The proposed features are efficient to compute due to the usage of the integral image and the vectorized operations. The proposed detector shows decent performance with respect to speed and accuracy.