High-accuracy real-time pedestrian detection system using 2D and 3D features

We present a real time stereo-vision pedestrian detector implementation with a very high accuracy, the 2D component of which attains 99% recall with less than 10(-6) false positives per window on the INRIA persons dataset. We utilize a sequence of classifiers which use different features, beginning with Haar-like features and a Haar-like feature implementation adapted to disparity images, and performing a final verification with Histogram-of-Oriented Gradient (HOG) features. We present a 2D Haar-like feature implementation that utilizes 2x2 kernel filters at multiple scales rather than integral images, and combines a quickly trained preliminary adaBoost classifier with a more accurate SVM classifier. We also show how these Haar-like features may be computed from a partially incomplete stereo disparity image in order to make use of 3-dimensional data. Finally, we discuss how these features, along with the HOG features, are computed rapidly and how the classifiers are combined in such a way as to enable real-time implementation with higher detection rates and lower false positive rates than typical systems. Our overall detector is a practical combination of speed and detection performance, operating on 544x409 image (10,425 windows) at a frame rate of 10-20fps, depending on scene complexity. The detector's overall false positive rate is less than 10(-6), corresponding to about one false positive every 10-60s when testing on our non-training data. Additionally, the detector has shown usefulness for detecting other object types, and has been implemented for traffic cones, telephone poles, and vehicles.