Robust Adaptive Classifier Grids for Object Detection from Static Cameras

In this work we present a robust object detection system for static cameras, which is suitable for real-time applications. Thus, the system has to cope with changes of environmental conditions, which is realized by adaptive on-line learning a scene specific classifier. In particular, we apply the ideas of grid-based classification, where each image patch corresponds to one classifier. Thus, the complexity of the detection task is reduced and a more compact and thus more efficient representation can be applied. The main contribution of this paper is to introduce three learning strategies to improve the performance of grid-based detectors: (a) pre-selecting features to assure a more efficient representation, (b) pre-training the positive representation, and (c) combining off-line and on-line learning. The experimental results on person and car detection show that these strategies significantly improve the overall performance of the detection system. In addition, a long-term experiment demonstrates that the proposed system is stable over time and can thus be applied for real-world tasks.