An aerial change detection system using multiple detector fusion and adaboost classification

Change detection in aerial imagery is challenging due to the effects of platform movements, atmospheric degradation and unpredictable photometric or thermal variations of the scene. We introduce an aerial change detection system built around multiple change detectors, each exploiting a different feature set, and fusion of their output in order to overcome the challenges. A supervised classifier screens the fused results for false detections. In particular, we have investigated three change detectors: Histogram of Oriented Gradients (HOG), Clustering-Based Anomaly Detection (CBCD), and Local Binary Similarity Pattern (LBSP). A set of change regions (blobs) are retained following fusion. Next, a set of features are extracted from each blob and fed into an AdaBoost classifier for true change acceptance test. Ours is a multithreaded system performing change detection in real time. Test results on a variety of EO and IR imagery demonstrate the efficacy our system in real-world applications.