Pedestrian Detection

You are in Home->Projects->Pedestrian Detection

Motivation

Automatic human activity recognition from video has recently attracted the attention of many researchers. It plays a critical role in surveillance systems that aim to know what the objects are, and what they are doing. The periodic nature of human motion has been widely used in gait recognition and related applications. The goal of this work is to classify an object as either a human or a vehicle based on its motion pattern.

Background

Among the many moving object classification methods, motion signature analysis is a simple and promising approach, especially for infrared and airborne video processing, which typically have low image contrast and small object size. Periodic motion signatures are robust low level clues in these situations.

Methodology

The main idea includes using a finite frequency set to probe the images of an object for its periodic signature and using the period and its strength for classification. A concise signal is derived from the periodic and symmetrical nature of human motion as an a priori reference. The method is efficient due to low computation cost. The period detection is transformed into a global-maximum location process. It works well for low contrast and small size targets where other methods have difficulties.

Results and Future Directions

A periodicity motion detection based object classification algorithm is reported. The method is simple, efficient, and robust to target size and frame rate. It transforms the complicated period detection into an easier global maximum location process. The choice of the probing by a priori reference signal within a finite frequency set enables accurate object classification even with a short video clip (2-3 seconds). Sensitivity analysis reveals that robust nature of the proposed method.

Publications and Awards

Jul 2004: Winning the "VIVID AWARD: Best Technical Performance Spiral II" with our team in UMD ( Larry Davis, Rama Chellappa, Wael, Qinfen, Atonio etc) and SRI.