Vehicle Detection in High-Resolution Aerial Images with Parallel RPN and Density-Assigner

Abstract

Vehicle detection in aerial images plays a significant role in many remote sensing applications such as city planning, road construction, and traffic control. However, detecting vehicles in aerial images remains challenging due to the existence of tiny objects, the scale variance within the same type of vehicle objects, and dense arrangement in some scenarios, such as parking lots. At present, many state-of-the-art object detectors cannot generate satisfactory results on vehicle detection in aerial images. The receptive field of the current detector is not fine enough to handle the slight scale variance. Moreover, the densely arranged vehicles will introduce ambiguous positive samples in label assignment and false predictions that cannot be deleted by NMS. To this end, we propose a two-stage framework for vehicle detection that better leverages the prior attribution knowledge of vehicles in aerial images. First of all, we design a Parallel RPN that exploits convolutional layers of different receptive fields to alleviate the scale variation problem. To tackle the densely arranged vehicles, we introduce a density-based sample assigner in the vehicle-intensive areas to reduce low-quality and occluded positive samples in the training process. In addition, a scale-based NMS is proposed to filter out redundant proposals hierarchically from different levels of the feature pyramid. Moreover, we construct two challenging vehicle detection datasets based on the AI-TOD and xView datasets which contain many tiny objects. Extensive experiments on these two datasets demonstrate the effectiveness of our proposed method.