Reduced training of convolutional neural networks for pedestrian detection

Pedestrian detection is a vision task with many practical applications in video surveillance, road safety, autonomous driving and military. However, it is much more difficult compared to the detection of other visual objects, because of the tremendous variations in the inner region as well as the outer shape of the pedestrian pattern. In this paper, we propose a pedestrian detection approach that uses convolutional neural network (CNN) to differentiate pedestrian and non-pedestrian patterns. Among several advantages, the CNN integrates feature extraction and classification into one single, fully adaptive structure. It can extract two-dimensional features at increasing scales, and it is relatively tolerant to geometric, local distortions in the image. Although the CNN has good generalization performance, training CNN classifier is time-comsuming. Therefore, we present an efficient training approach for CNN. Through the experiments, we show that it is possible to design networks in a fraction of time taken by the standard learning approach.