Low resolution pedestrian detection using light robust features and hierarchical system

The pedestrian detection is a popular research field in recent years, yet the low-resolution issue is rarely discussed for yielding detection accuracy for drivers. In this study, a hierarchical pedestrian detection system is proposed to cope with this issue. In which, two independent features, orientation and magnitude, are adopted as descriptors for pedestrians. Moreover, the proposed probability-based pedestrian mask pre-filtering (PPMPF) is utilized to initially filter out non-pedestrian regions meanwhile retaining most of the real pedestrians. In experimental results, the use of the two proposed features can provide superior performance than the former well-known histogram of oriented gradient (HOG; high accuracy) and the edgelet (high processing efficiency) simultaneously without carrying their lacks. Moreover, the PPMPF can also boost the processing efficiency by a factor of around 2.82 in contrast to the system without this pre-filtering strategy. Thus, the proposed method can be a very competitive candidate for intelligent surveillance applications.