• 
• 
• 
• 
• 
• 

Pedestrian Detection

MERL is developing new technology for automatically detecting pedestrians in video sequences.  Detecting and tracking pedestrians is important for surveillance purposes.  It can be used to sound an alarm if an intruder is in a restricted area or to aid in browsing hours of surveillance video by skipping to the next part of the video where a person was seen.  

Background & Objective:  The problem of detecting people in low resolution surveillance video is a fundamental problem for surveillance.  The problem is difficult because the pedestrian may be very small in the image, making the amount of information contained in the pixels small.  Furthermore, there may be background motion in the scene (such as trees waving in the breeze or a cloud shadow passing by) which makes motion detectors faulty.  To overcome these problems, we learn a model that encompasses both the appearance (pixels) and the motion of pedestrians.  The pedestrian model is fast to evaluate which makes it feasible to search for pedestrians across the frames of a video.

Technical Discussion:  To detect pedestrians in a video sequence, we learn a foreground model of the motion and appearance of pedestrians from example video sequences.  The pedestrian detector builds on earlier face detection work at MERL.  The earlier work is extended by using motion information as well as appearance information.  This is in contrast to most prior work which attempts to build a model of the background.  In our case, no background model is used.  The detector learned uses a set of simple motion and appearance features.  The appearance features are simple rectangle features acting on a single frame of the video.  The motion features are simple rectangle features acting on the difference image between successive frames of the video.  The optimal set of motion and appearance features are learned from a large library of possible features using the AdaBoost learning algorithm.

Publications:
Viola, P.; Jones, M.J.; Snow, D., "Detecting Pedestrians Using Patterns of Motion and Appearance", IEEE International Conference on Computer Vision (ICCV), Vol. 2, pp. 734-741, October 2003 (IEEE Xplore, TR2003-090)

Technical Reports:
TR2005-065	Ensemble Tracking

Technology Area:  Computer Vision

Modification Date:  June 13, 2008