Normalized Autobinomial Markov Channels For Pedestrian Detection

This paper brings significant contributions to the field of pedestrian detection by learning probabilistic dependencies and contextual information that draw special attention to the human body characteristics and silhouette shapes and play down other irrelevant features. More precisely, we introduce the NAMC (Normalized Autobinomial Markov Channels) and study the efficiency of different configurations of cliques, providing a detailed experimental evaluation. Our proposed features outperform most of the solutions that have laid the foundations of pedestrian detection [2, 17]. Moreover, if we combine our novel features with gradient-based descriptors [15] and apply an efficient local decorrelation algorithm [30] to each channel, our results outperform the majority of the state-of-the-art solutions currently present in the Caltech Pedestrian Detection Benchmark [13]. We focus on a thorough analysis of the proposed feature model using the INRIA Pedestrian Dataset [10] as a benchmark to evaluate various parameter settings.