Design of coupled strong classifiers in AdaBoost framework and its application to pedestrian detection

Abstract

In the AdaBoost framework, a strong classifier consists of weak classifiers connected sequentially. Usually the detection performance of the strong classifier can be improved increasing the number of weak classifiers used, but the improvement is asymptotic. To achieve further improvement we propose coupled strong classifiers (CSCs) which consist of multiple strong classifiers connected in parallel. Complementarity between the classifiers is considered for reducing intra- and inter-classifier correlations of exponential loss of weak classifiers in the training phase, and dynamic programming is used during the testing phase to compute efficiently the final object score for the coupled classifiers. In addition to CSC concept, we also propose using Aggregated Channel Comparison Features (ACCFs) that take the difference of feature values of Aggregated Channel Features (ACFs), enabling significant performance improvement. To show the effectiveness of our CSC concept, we apply our algorithm to pedestrian detection. Experiments are conducted using four well-known benchmark datasets based on ACFs, ACCFs, and Locally Decorrelated Channel Features (LDCFs). The experimental results show that our CSCs give better performance than the conventional single strong classifier for all cases of ACFs, ACCFs, and LDCFs. Especially our CSCs combined with ACCFs improve the detection performance significantly over ACF detector, and its performance is comparable to those of the state-of-the-art algorithms while using the simple ACF-based features.