Rotated Fusion Network (RFN) Algorithm for Aerial Vehicle Detection

The improvement of intelligent transportation systems requires high-precision vehicle detection under full-weather condition. From the view of unmanned aerial vehicle (UAV), vehicle information in the global scope can be monitored effectively. At present, most of vehicle datasets are monomodal or unregistered multimodal, failing to be trained to detect vehicles under the harsh environment. In this paper, a visible and infrared aerial vehicle dataset, named VIA, has been proposed, which is highly temporal registered and covers a variety of scenarios. In addition, a visible and infrared rotation fusion detection framework based on YOLOv3, named RFN, has been proposed. The framework uses a dual-path feature extraction network to extract features separately and uses rotated bounding boxes to locate vehicles, solving the problem of confusion and redundancy of vehicle information caused by horizontal bounding box detection. Besides, the effectiveness of Oriented Response Network (ORN) and Asymmetric Convolution Blocks (ACB) modules for feature enhancement is explored. The promising results have been obtained in the VIA dataset by this method.