Vehicle Detection in Aerial Images: A Survey

In a variety of computer vision-based applications, one of the most crucial jobs is the recognition of vehicles from unmanned aerial vehicle (UAV) imagery. High accuracy and speed were required to do this critical assignment. Aerial image properties and the technology being utilized, such as diverse automobile sizes and their orientations, densities, small datasets, and inference speed, make this a highly difficult assignment. Numerous approaches based on hand-crafted feature design and deep learning have been put out in the literature as solutions to these issues in recent years. Techniques based on hand-crafted features and shallow learning have limited applicability to other complicated scenarios. On the other hand, Due to deep learning's robust learning capability, vehicle detection algorithms produced superior outcomes. In this article, we reviewed the hand-crafted feature design-based and data-driven vehicle detection algorithms from UAV imagery. We start by presenting various hand-crafted design-based vehicle detection algorithms like Histogram of oriented gradients (HOG), Scale Invariant Feature Transform (SIFT), and Local Binary Pattern (LBP). After that, we reviewed one-stage and two-stage vehicle detection series based on deep learning. One-stage includes state-of-art algorithms like YOLO and SSD, and the two-stage series includes object detection models like R-CNN, Fast R-CNN, Faster-RCNN, etc. We also focused on investigating the various public aerial image datasets and their related work. In the end, we have also shed some insight into the conclusion of this paper.