Parts-based object recognition seeded by frequency-tuned saliency for child detection in active safety

This paper proposes a novel system for automatically detecting children from a color monocular back-up camera, as part of a back-up warning device in passenger vehicles. We presented the use of an attentional mechansim that focuses compute-intensive bounding-box classifiers on a subset of all possible bounding-box solutions to enable real-time performance of 248ms per frame with negligible reduction in performance. The attentional mechanism called Attention to Children which consists of a window generation and verification cascade of based on Frequency-Tuned Saliency, Variational-Optical-Flow Obstacle Detection and finally a parts-based classifier. We also presented a method of reducing much of the cascade classifier evaluations by judicious sampling of the bounding-box solution space. The result is a reduction in the number of windows evaluated down to 439 from more than 12K windows in traditional sliding window techniques, a 97% reduction in the number of windows. The verification stages leading up to the parts-based classifier further reduces the number of windows to half. Together with a parallel processing and pipelining, the final processing time was 248ms per frame.