Context-aware pedestrian detection especially for small-sized instances with Deconvolution Integrated Faster RCNN (DIF R-CNN)

Pedestrian detection is a canonical problem in computer vision. Motivated by the observation that the major bottleneck of pedestrian detection lies on the different scales of pedestrian instances in images, our effort is focused on improving the detection rate, especially for small-sized pedestrians who are relatively far from the camera. In this paper, we introduce a novel context-aware pedestrian detection method by developing the Deconvolution Integrated Faster R-CNN (DIF R-CNN), in which we integrate a deconvolutional module to bring additional context information which is helpful to improve the detection accuracy for small-sized pedestrian instances. Furthermore, the state-of-the-art CNN-based model (Inception-ResNet) is exploited to provide a rich and discriminative hierarchy of feature representations. With these enhancements, a new synthetic feature map can be generated with a higher resolution and more semantic information. Additionally, atrous convolution is adopted to enlarge the receptive field of the synthetic feature map. Extensive evaluations on two challenging pedestrian detection datasets demonstrate the effectiveness of the proposed DIF R-CNN. Our new approach performs 12.29% better for detecting small-sized pedestrians (those below 50 pixels in bounding-box height) and 6.87% better for detecting all case pedestrians of the Caltech benchmark than the state-of-the-art method. For aerial-view small-sized pedestrian detection, our method achieve 8.9% better performance when compared to the baseline method on the Okutama human-action dataset.