Moving vehicle detection in aerial infrared image sequences via fast image registration and improved YOLOv3 network

Owing to its powerful adaptability and robustness to the weak light, the infrared camera equipped on the unmanned aerial vehicle is more and more applied for aerial photography. Nowadays, how to make full use of the aerial infrared image sequences for the moving vehicle detection has attracted widespread attention. However, due to the low-resolution, low contrast, and few texture features of the infrared image, it is extremely difficult to detect the moving vehicles. In our work, an effective and efficient moving vehicle detection approach in the aerial infrared image sequences is proposed via fast image registration and You Only Look Once Version 3 (YOLOv3) network. First, to compensate for the motion of the aerial infrared camera, a fast infrared image registration method is put forward. To improve the accuracy and efficiency of image registration, we construct a multi-screening based mechanism for screening out the incorrect and redundant feature points. For feature description, the low-level and high-level descriptors are combined to further improve the registration accuracy. Then, the 2-frame difference and image masking are introduced to acquire the frame mask images, where only the region-of-interest is reserved, and the remaining regions are masked. Next, we construct a new structure of an improved YOLOv3 network with only 23 layers. Due to the insufficiency of the infrared vehicle samples, transfer learning is introduced to train the improved YOLOv3 network. Finally, the proposed approach is evaluated on the Defence Advanced Research Projects Agency (DARPA) Video verification of Identify (VIVID) and Northwest Polytechnical University (NPU) data sets. Experiments and comprehensive analyses show that the proposed approach can achieve satisfactory and competitive moving vehicle detection results.