Two Co-occurrence Histogram Features Using Gradient Orientations and Local Binary Patterns for Pedestrian Detection

Pedestrian detection plays important roles in various applications such as automobile driving assistance and surveillance camera system. The co-occurrence histograms of oriented gradients (CoHOG) feature descriptor showed good performance since thirty co-occurrences at each pixel position represent various spatial characteristics of object shapes. Though extraction of co-occurrence histogram features is computationally demanding, there is an application-specific integrated circuit (ASIC) to accelerate the computation. The hardware accelerator enables CoHOG to be used in real-time applications. In this paper, we propose the use of two co-occurrence histogram features describing different aspects of object shapes to improve accuracy of pedestrian detection. One feature is CoHOG and the other is co-occurrence histograms of local binary patterns (CoHLBP). CoHLBP assigns each pixel into eight categories by comparing a center pixel's value and its three neighbors' values, and then co-occurrence histograms are calculated in the same way as for CoHOG. Since the number of local binary patterns is the same as the number of quantized orientations used in CoHOG, the CoHOG hardware accelerator can be used for CoHLBP calculation. The experimental results using the benchmark NICTA pedestrian dataset show that the proposed method reduces the false positive rate to less than one-quarter of that of CoHOG.