About AMP Lab Projects Downloads Publications People Links Project - Real-time Pedestrian Detection using Eigenflow |
| Motivation and Goal |
|
Every day there are incidents of pedestrians being hit by a car. Majority of the accidents occur either at the pedestrian crossings or while reversing the car. In either case, it’s the inability of the driver to perceive the pedestrians that causes the trouble. The rear mirrors don’t always provide a full view of the scene behind the car and this situation becomes worse for kids and babies, owing to their smaller size. |  |
 |
In this project, we aim to develop a robust pedestrian detection system for GM cars that can alert the driver of the presence of the pedestrians behind the car and thus avert a possible collision. A camera mounted at the back of the car would capture the scene behind and an algorithm, operating on the video, would detect any moving pedestrians. |
Top of this page
Overview
We propose a novel learning algorithm to detect moving pedestrians from a stationary camera. The algorithm learns a discriminative model based on eigenflow, i.e. the eigen vectors derived from applying Principal Component Analysis to the optical flow of moving objects, to differentiate between human motion patterns from other kind of motions like cars etc. The learned model is a cascade of Adaboost classifiers of increasing complexity, with eigenflow vectors as weak classifiers.
Detection Results
Unlike some recent attempts to use motion for pedestrian detection, this system performs this task in real-time @ 10fps on a Pentium M 1.6 GHz machine. The system is also robust to small camera jitter and changes. Moreover, we are able to detect moving children using the same system even though the training data is mainly composed of adult pedestrians.
|
- D. Goel and T. Chen, "Real-time Pedestrian Detection using Eigenflow," IEEE International Conference on Image Processing, 2007 (Submitted).
Any suggestions or comments are welcome. Please send them to Dhiraj Goel.
|