An unsupervised, online learning framework for moving object detection

Abstract

Object detection with a learned classifier has been applied successfully to difficult tasks such as detecting faces and pedestrians. Systems using this approach usually learn the classifier offline with manually labeled training data. We present a framework that learns the classifier online with automatically labeled data for the specific case of detecting moving objects from video. Motion information is used to automatically label training examples collected directly from the live detection task video. An online learner based on the Winnow algorithm incrementally trains a taskspecific classifier with these examples. Since learning occurs online and without manual help, it can continue in parallel with detection and adapt the classifier over time. The framework is demonstrated on a person detection task for an office corridor scene. In this task, we use background subtraction to automatically label training examples. After the initial manual effort of implementing the labeling method, the framework runs by itself on the scene video stream to gradually train an accurate detector.

Automatic adaptation of a generic pedestrian detector to a specific traffic scene

“…They controls how aggressive the automatic training process is. Similar parameters also exist in other approaches of automatically training scene specific detectors [10,11,12,13,14,15,16]. Our approach has robustness to these parameters within certain range.…”

“…Positive and negative examples for re-training are automatically selected according to the number of foreground pixels within the detection windows. A similar strategy was used in [13]. [13] used Haar features + Boosting.…”

“…A similar strategy was used in [13]. [13] used Haar features + Boosting. To make the comparison consistent, we still use SVM + HOG.…”

“…By calculating the frame difference as 0.5(|I t − I t−50 | + I t − I t+50 |), moving pixels inside a detection window are thresholded and counted. Similar to other self-training [15,13] or co-training [11,10,16,21] frameworks, the confident positive examples are found by thresholding L p (z) > L 0 . The larger the threshold is, the more conservative the strategy of selecting examples is.…”

“…Since the detector itself was the labeler and not reliable, the selected examples were not informative and likely to have the wrong labels, which made the detector drift. Nair et al [13] used background subtraction results to label training examples for an appearance-based pedestrian detector. The accuracy of the background subtraction labeler was low and it introduced biased labeling which misled the learning of the detector.…”