Improving pedestrian detection using light convolutional neural network

Convolution Neural Networks (CNNs) have been used widely in computer vision field and have brought enormous gain to image classification state-of-the-art. Recently, authors tend to increase the CNN depth (using more layers) to reach higher accuracy. However, such approach results in a waste of computational time and space. In this paper, we propose a light convolutional neural network method for pedestrian image classification. The network consists of three convolutional layers, three max-pooling layers and one fully connected layer. By designing a small network, we aim to achieve high classification performance while saving computational costs. Additionally, we conduct an in-depth study on the characteristics of our CNN features and use them separately with an SVM classifier. Experiments performed on the Inria benchmark dataset show the effectiveness and the high accuracy achieved by our CNN in both feature extraction and classification scenarios.