R-SSD: refined single shot multibox detector for pedestrian detection

Pedestrian detection is a critical task in the field of computer vision, and it has made considerable progress with the help of Convnets. However, a persistent crucial problem is that small-scale pedestrians are notoriously difficult to detect because of the introduction of weak contrast and blurred boundaries in real-world scenarios. In this paper, we present a simple and compact detection method for detecting multi-scale pedestrians, which is especially suitable for detecting small-scale pedestrians that are not easily recognized in images or videos. We first interpret convolutional neural network (CNN) channel features, explore the detection performance of different feature fusion methods, and propose a novel two-level feature fusion strategy specially designed for small-scale pedestrians. Moreover, a sub-network named “prediction module” is injected into the framework to improve the general performance without any bells and whistles. In addition, we propose an adaptive loss that adds an adaptive adjustment coefficient to the Smooth L1 loss function to enhance its robustness to pedestrian detection tasks. Using these methods synthetically, we achieve state-of-the-art detection performance on the Caltech pedestrian dataset under three evaluation protocols; particularly, the performance of small-scale pedestrians under “Far” evaluation setting is improved (miss rate decreases from 70.97% to 60.09%). Further, the proposed method achieves a competitive speed-accuracy trade-off with 0.31 second per image of 1024×2048 pixels on the CityPersons dataset.