Abstract
Pedestrian detection required for numerous practical applications is one of the most challenging problems. Recently, accurate schemes for pedestrian detection have been proposed and achieved acceptable detection accuracy for several practical applications. However, the increase of computational complexity becomes a significant problem, because most recent schemes adopt sophisticated feature descriptors and advanced machine learning algorithms. To solve this problem, computational complexity reduction schemes and parallel implementations using hardware and multi-core processors are proposed, but the processing speed still remains insufficient for real-time computation. Considering this background, we propose an acceleration scheme that can be combined with existing schemes, and show experimental results using CoHOG-based pedestrian detection. In the proposed scheme, the number of sampling is reduced by efficient sampling based on the probability distribution computed from the results of sliding window detection at reference images. Experimental results using INRIA data set show that the proposed scheme can compute about 2.5 times as fast as the original implementation without any degradation of detecting accuracy where false positive per image (FPPI) is adopted as a measure.
