보행자 검출기의 직렬연결을 통한 성능 개선

With advancement of Convolutional Neural Network (ConvNet), recent developments in pedestrian detection are gaining much improvements. ConvNet is trained with massive pedestrian dataset and the features from the trained ConvNet can handle variations in pedestrian such as viewpoint and pose. But the classification performance of ConvNet depends on how the dataset is constructed. In this paper, we propose to train two ConvNets with different composition of dataset and to detect pedestrian by cascading these networks. Experiments in INRIA pedestrian dataset demonstrate the effectiveness of proposed method.