Efficient scan-window based object detection using GPGPU

We describe an efficient design for scan-window based object detectors using a general purpose graphics hardware computing (GPGPU) framework. While the design is particularly applied to built a pedestrian detector that uses histogram of oriented gradient (HOG) features and the support vector machine (SVM) classifiers, the methodology we use is generic and can be applied to other objects, using different features and classifiers. The GPGPU paradigm is utilized for feature extraction and classification, so that the scan windows can be processed in parallel. We further propose to precompute and cache all the histograms in advance, instead of using integral images, which greatly lowers the computation cost. A multi-scale reduce strategy is employed to save expensive CPU-GPU data transfers. Experimental results show that our implementation achieves a more-than-ten-times speed up with no loss on detection rates.