Digital

Pedestrian detection in crowded scenes with the histogram of gradients principle

Publikation aus Digital

Sidla O., Rosner M., Lipetski Y.

SPIE - Intelligent Robots and Computer Vision XXIV: Algorithms, Techniques and Active Vision; Boston, USA Oct 2006 , 2006

Abstract:

This paper describes a close to real-time scale invariant implementation
 of a pedestrian detector system which is based on the Histogram of
 Oriented Gradients (HOG) principle. Salient HOG features are first
 selected from a manually created very large database of samples with
 an evolutionary optimization procedure that directly trains a polynomial
 Support Vector Machine (SVM). Real-time operation is achieved by
 a cascaded 2-step classifier which uses first a very fast linear
 SVM (with the same features as the polynomial SVM) to reject most
 of the irrelevant detections and then computes the decision function
 with a polynomial SVM on the remaining set of candidate detections.
 Scale invariance is achieved by running the detector of constant
 size on scaled versions of the original input images and by clustering
 the results over all resolutions. The pedestrian detection system
 has been implemented in two versions: i) fully body detection, and
 ii) upper body only detection. The latter is especially suited for
 very busy and crowded scenarios. On a state-of-the-art PC it is able
 to run at a frequency of 8 - 20 frames/sec.