Unmanned aerial vehicles (UAVs) play an important role in conducting automatic patrol inspections of cities, which can ensure the safety of urban residents’ life and property and the normal operation of cities. However, during the inspection process, problems may arise. For example, numerous small objects in UAV images are difficult to detect, objects in UAV images are severely occluded and requirements for real-time performances are posed. To address these issues, we first propose a real-time object detection network (RTD-Net) for UAV images. Besides, to deal with the lack of visual features of small objects, we design a feature fusion module (FFM) to interact and fuse features at different levels and improve the feature expression ability of small objects. To achieve real-time detection, we design a lightweight feature extraction module (LEM) to build the backbone network to control the calculation quantity and parameters. To solve the issue of discontinuous features of occluded objects, an efficient convolutional Transformer block (ECTB) based convolutional multi-head self-attention (CMHSA) is designed to improve the recognition ability of occluded objects by extracting the context information of objects. Compared with multi-head self-attention (MHSA) in the traditional transformer, CMHSA uses convolutional projection to replace the position-linear projection, which can reduce a large amount of calculation without performance loss. Finally, an attention prediction head (APH) is designed based on the attention mechanism to improve the ability of the model to extract attention regions in complex scenarios. The proposed method reaches a detection accuracy of 86.4% mAP in our UAV image dataset. In addition, it achieves a detection accuracy of 86.0%mAP and a detection speed of 33.4 FPS in NVIDIA Jeston TX2 embedded device.

Unmanned aerial vehicles (UAVs) have been widely used in postdisaster search and rescue operations, object tracking, and other tasks. Therefore, the autonomous perception of UAVs based on computer vision has become a research hotspot in recent years. However, UAV images include dense objects, small objects, and arbitrary object directions, which bring about significant challenges to existing object detection methods. To alleviate these issues, we propose a global-local feature enhanced network (GLF-Net). Considering the difficulty of processing UAV images with complex scenes and dense objects, we designed a backbone based on an involution and self-attention that can extract effective features from complex objects. A multiscale feature fusion module is also proposed to address the presence of numerous small objects in UAV images through multiscale object detection and feature fusion. To accurately detect rotated objects, a rotated regional proposal network was designed based on the midpoint offset representation, which can apply a rotated box to determine the real direction and contour of an object. GLF-Net achieves a state-of-the-art detection accuracy [86.52% mean average precision (mAP)] on our created rotated object detection UAV (RO-UAV) dataset, while achieving 96.95% and 97% mAP on the public datasets high resolution ship collections 2016 (HRSC2016) and the University of Chinese Academy of Sciences High Resolution Aerial Object Detection Dataset (UCAS-AOD), respectively. The experimental results demonstrate that our method achieves a high detection accuracy and generalization, which can meet the practical requirements of UAVs under various complex scenarios.

Many accidents happen under shunting mode when the speed of a train is below 45 km/h. In this mode, train attendants observe the railway condition ahead using the traditional manual method and tell the observation results to the driver in order to avoid danger. To address this problem, an automatic object detection system based on convolutional neural network (CNN) is proposed to detect objects ahead in shunting mode, which is called Feature Fusion Refine neural network (FR-Net). It consists of three connected modules, i.e., the depthwise-pointwise convolution, the coarse detection module, and the object detection module. Depth-wise-pointwise convolutions are used to improve the detection in real time. The coarse detection module coarsely refine the locations and sizes of prior anchors to provide better initialization for the subsequent module and also reduces search space for the classification, whereas the object detection module aims to regress accurate object locations and predict the class labels for the prior anchors. The experimental results on the railway traffic dataset show that FR-Net achieves 0.8953 mAP with 72.3 FPS performance on a machine with a GeForce GTX1080Ti with the input size of 320 × 320 pixels. The results imply that FR-Net takes a good tradeoff both on effectiveness and real time performance. The proposed method can meet the needs of practical application in shunting mode.

With the high growth rates of railway transportation, it is extremely important to detect railway obstacles ahead of the train to ensure safety. Manual and traditional feature-extraction methods have been utilized in this scenario. There are also deep learning-based railway object detection approaches. However, in the case of a complex railway scene, these object detection approaches are either inefficient or have insufficient accuracy, particularly for small objects. To address this issue, we propose a feature-enhanced single-shot detector (FE-SSD). The proposed method inherits a prior detection module of RON and a feature transfer block of FB-Net. It also employs a novel receptive field-enhancement module. Through the integration of these three modules, the feature discrimination and robustness are significantly enhanced. Experimental results for a railway traffic dataset built by our team indicated that the proposed approach is superior to other SSD-derived models, particularly for small-object detection, while achieving real-time performance close to that of the SSD. The proposed method achieved a mean average precision of 0.895 and a frame rate of 38 frames per second on a railway traffic dataset with an input size of $320\times320$ pixels. The experimental results indicate that the proposed method can be used for real-world railway object detection.

Obstacles in front of a train pose a significant threat to traffic safety, and many accidents happen under shunting mode when the speed of a train is below 45 km/h. The existing track object–detection algorithms encounter difficulty in balancing the detection precision and speed in shunting mode. Additionally, their accuracy is insufficient, particularly for small objects in complex environments. To address these problems, we propose a stable lightweight feature extraction and adaptive feature fusion network for real-time detection of obstacles in railway traffic scenarios to ensure driving safety. The proposed network consists of three modules. The stable bottom feature extraction module reduces the computational load and extracts more image information stably. The lightweight feature extraction module improves feature extraction using a simple and effective network. The enhanced adaptive feature fusion module fuses the image and original features, improving the multiscale detection accuracy under complex environments, particularly in the case of small objects. With a default input size of 416 $\times 416$ pixels (px), the proposed method achieves a detection speed of 81 FPS and a mean average precision of 94.75% for the railway traffic dataset as well as a detection speed of 78 FPS (26 FPS faster and 0.47% higher than those of YOLOv4, respectively) and a mean average precision of 42.5% for MS COCO. This indicates its potential for real-world railway object detection and other multi-target detection tasks. Additionally, the experimental results based on PASCAL VOC2007 and VOC2012 indicate that the proposed approach is considerably better than the state-of-the-art models.

With the growing popularity of civilian unmanned aerial vehicles (UAVs), unauthorized flights are on the rise accordingly. Therefore, it is critical to detect low-altitude UAVs for protecting personal privacy and public safety. Though substantial progress has been made in UAV detection, the existing detection methods still have problems in balancing the detection accuracy, the model size, and the detection speed. To address these limitations, this article proposes a novel deep learning method named convolution–transformer network (CT-Net). First, the attention-enhanced transformer block (AETB), which builds a feature-enhanced multihead self-attention (FEMSA), is introduced into the backbone of the network to improve the feature extraction ability of the model. Then, a lightweight bottleneck module (LBM) is utilized to control the computation load and reduce the parameters. Finally, we present a directional feature fusion structure (DFFS) to improve the accuracy of detection when processing multiscale objects, especially small-size objects. The proposed scheme achieves 0.966 mAP with the input size of 640 $\times640$ pixels on our low-altitude small-object dataset, and the performance is superior to YOLOv5. Furthermore, the experimental result on MS COCO shows that the CT-Net can outperform the current state-of-the-art detectors by a large margin. Thus, it is feasible to apply CT-Net to low-altitude small-object detection according to the experimental results.