G2P: A new descriptor for pedestrian detection

Pedestrian detection plays an important role in intelligent vehicle applications. Since its birth 12 years ago, the Histogram-Of-Gradient (HOG) descriptor has become a popular descriptor for pedestrian detection, thanks to its effectiveness in capturing implicit human characteristics. Besides its original instantiation, the HOG also reflects a general methodology of constructing descriptors based on histograms of gradients of certain image sub-blocks. Following this general methodology, a number of HOG-style descriptors have been reported in literature. Three contributions are made in this work. First, a general model called Descriptor Generation Model (DGM) is proposed, which can be used to systematically construct a wide range of HOG-style descriptors for pedestrian detection. Second, based on the DGM, a pedestrian detection experimental framework (PDEF) is introduced to find the optimal HOG-style descriptor. In the PDEF, the performance of each descriptor can be evaluated. At last, the genetic algorithm is employed to search the optimal (or semi-optimal) HOG-style descriptor in the descriptor space. And a new descriptor named Second-order Gradient for Pedestrian detection (G2P) is presented. Experimental results demonstrate the advantage of the G2P descriptor over the standard HOG descriptor with ETH, CVC-02-system, NITCA and KITTI dataset, which also reflects the effectiveness of the DGM-based PDEF in finding better descriptors for pedestrian detection.