Vehicle detection in aerial images is an important and challenging task in the field of remote sensing. Recently, deep learning technologies have yielded superior performance for object detection in remote sensing images. However, the detection results of the existing methods are horizontal bounding boxes that ignore vehicle orientations, thereby having limited applicability in scenes with dense vehicles or clutter backgrounds. In this article, we propose a one-stage, anchor-free detection approach to detect arbitrarily oriented vehicles in high-resolution aerial images. The vehicle detection task is transformed into a multitask learning problem by directly predicting high-level vehicle features via a fully convolutional network. That is, a classification subtask is created to look for vehicle central points and three regression subtasks are created to predict vehicle orientations, scales, and offsets of vehicle central points. First, coarse and fine feature maps outputted from different stages of a residual network are concatenated together by a feature pyramid fusion strategy. Upon the concatenated features, four convolutional layers are attached in parallel to predict high-level vehicle features. During training, task uncertainty learned from the training data is used to weight loss function in the multitask learning setting. For inferencing, oriented bounding boxes are generated using the predicted vehicle features, and oriented nonmaximum suppression (NMS) postprocessing is used to reduce redundant results. Experiments on two public aerial image data sets have shown the effectiveness of the proposed approach.

Vehicle detection in satellite videos is crucial for large-scale traffic monitoring and urban management, yet remains challenging due to the extremely small object size and limited appearance features. Existing methods generally focus on extracting spatiotemporal information at the local pixel level while neglecting the long-range dependencies between vehicle instances. This insufficient spatiotemporal feature aggregation limits the accuracy of detection. To address this issue, we propose a coarse-to-fine multi-level spatiotemporal attention network (MLSTA-Net) for detecting tiny vehicles in satellite videos. Specifically, a pixel-level spatiotemporal attention (PiSTA) module is introduced, which leverages motion priors to guide the aggregation of spatiotemporal features at the pixel level, thereby enhancing the fine-grained representation of targets and generating coarse vehicle detection results. Subsequently, an instance-level spatiotemporal attention (InSTA) module is developed to refine initial detections by modeling the spatial relationships of instances within a single frame and the temporal consistency of instances across multiple frames. Experiments on VISO and SAT-MTB datasets demonstrate that the MLSTA-Net method achieves superior performance over state-of-the-art object detection approaches.

Geographic height information describes the vertical spatial structure of the city and serves as important foundational data for urban management. Obtaining height information from single-view remote sensing images is a relatively low-cost and convenient approach. However, there exist several bottlenecks in the current methods for inferring height information from monocular remote sensing images, such as difficulties in learning three-dimensional semantic information and accurately fitting the height morphology in various local scenes. In this study, we address these challenges by proposing a morphological priors guided network, termed MPG-Net, for accurate height estimation from single-view remote sensing images. First, considering the semantic morphological priors, we propose to explicitly enhance the three-dimensional visual cues (e.g., co-occurrence relationship between shadow-buildings and shadow-trees) and simultaneously design a semantic booster composed of two-stream network with multi-level cross-stream attention fusion mechanism to facilitate the three-dimensional feature learning for monocular height estimation. Second, taking into account the height distribution priors, we propose a scalable bins module (ScaBins) that can create fully adaptive bins within a flexible height range for each input image, leading a more accurate delineation of height distribution pattern. The proposed MPG-Net is comprehensively evaluated on two datasets of different scenes (i.e., ISPRS Vaihingen and Potsdam datasets). Results indicate that the proposed MPG-Net significantly outperforms the existing methods, with lowest root mean square error of 1.613 m and 1.947 m on Vaihingen and Potsdam, respectively. Furthermore, extensive ablation studies demonstrate the contribution of each designed component in the proposed method.