Human victim detection in an urban search and rescue scenario is challenging owing to the articulated nature and unpredictable position of the human body. This study investigates the effects of using an ensemble of classifiers (AdaBoost, k-NN and SVM) with a set of different feature types (HOG and SURF) on the human victim detection problem. The classifier ensemble uses both majority voting and a decision rule based on classification history to determine the outcome. A training dataset of 1590 simulated disaster images acquired for this study is used for training and the proposed approaches are evaluated via k-fold cross validation and through tests conducted on video data. The novelty of our approach lies in the incremental learning component that acquires domain knowledge and trains in parallel without interrupting the ongoing classification process. The system achieves over 69% accuracy in detecting human victims in images of a simulated disaster scenario.
展开▼
