Instance Implant-Aided Non-uniformly Cropping for Person Detection in Aerial Images

Human inspection of aerial-based scenarios is a time-consuming, labor-intensive, and error-prone task. However, current aerial-vision-based object detection techniques have not adequately improved the detection performance of persons with sparse and weak features. In this article, a novel approach is proposed to automatically detect persons in aerial images. Specifically, numerous and multiscale instances are first implanted on fresh background images with a copy-paste mechanism to collect rich training data. Then, cropping high-resolution images into irregular sizes with density maps in the training stage. Lastly, hyper-inferencing on the raw and sliced blocks of the image by merging these overlap results to find persons. Experiments were run on the original Heridal datasets against eight approaches of popular object detectors. The metrics of precision, AP(test)@0.5, and AP(test)@0.5:0.95 achieved 0.9277 (+2.77%), 0.8012, and 0.547, respectively. Meanwhile, our proposed method is being evaluated on the modified Heridal and a number of publicly available datasets to make a more robust person detector in aerial images.