Pedestrian detection with convolutional neural networks

Abstract

This paper presents a novel pedestrian detection method based on the use of a convolutional neural network (CNN) classifier. Our method achieves high accuracy by automatically optimizing the feature representation to the detection task and regularizing the neural network. We evaluate the proposed method on a difficult database containing pedestrians in a city environment with no restrictions on pose, action, background and lighting conditions. The false positive rate (FPR) of the proposed CNN classifier is less than 1/5-th of the FPR of a support vector machine (SVM) classifier using Haar-wavelet features when the detection rate is 90%. The accuracy of the SVM classifier using the features learnt by the CNN is equivalent to the accuracy of the CNN, confirming the importance of automatically optimized features. The computational demand of the CNN classifier is, however, more than an order of magnitude lower than that of the SVM, irrespective of the type of features used.