Walker Detection in Outdoor Scenes Using Spatial-Temporal Feature Analysis of Variation Regions

摘要

We propose an image processing algorithm to detect walkers in outdoor images containing movement in the background such as swaying trees or irregular reflections from water. Previously proposed methods to detect walkers focus on the fact that locally distributed movement is reinforced along a spatial and temporal path marked out by an object such as a human that moves at approximately a constant speed and in a straight line while other background movement tends to cancel itself out. However, false alarms can occur with methods that are based on the time-averaged strength of movement due to mistaken associations between moving objects and the fact that temporary movement in high-contrast areas can significantly raise the average value. In our method we use features based on a spatially averaged strength of movement and the uniformity of movement over time in addition to the time-averaged strength of movement, attempting to distinguish walkers from background movement within a feature space made up of these three features. In experiments with real images we have confirmed that the proposed method can reduce the rate of both false positives and false negatives compared to a method that focuses only on the time-averaged strength of movement. In addition, we have constructed an on-line system to implement the proposed method and the results of evaluation experiments conducted in a real-life environment over a period of two weeks show that we were able to obtain a level of performances in which the failure rate was less than 1 percent and false alarms occurred less than three times per day.