Visual on-line learning in distributed camera networks

Automatic detection of persons is an important application in visual surveillance. In general, state-of-the-art systems have two main disadvantages: First, usually a general detector has to be learned that is applicable to a wide range of scenes. Thus, the training is time-consuming and requires a huge amount of labeled data. Second, the data is usually processed centralized, which leads to a huge network traffic. Thus, the goal of this paper is to overcome these problems, which is realized by a person detection system, that is based on distributed smart cameras (DSCs). Assuming that we have a large number of cameras with partly overlapping views, the main idea is to reduce the model complexity of the detector by training a specific detector for each camera. These detectors are initialized by a pre-trained classifier, that is then adapted for a specific camera by co-training. In particular, for co-training we apply an on-line learning method (i.e., boosting for feature selection), where the information exchange is realized via mapping the overlapping views onto each other by using a homography. Thus, we have a compact scene-dependent representation, which allows to train and to evaluate the classifiers on an embedded device. Moreover, since the information transfer is reduced to exchanging positions the required network-traffic is minimal. The power of the approach is demonstrated in various experiments on different publicly available data sets. In fact, we show that on-line learning and applying DSCs can benefit from each other.