Aiming at the problems of low detection accuracy and large computing resource consumption of existing Unmanned Aerial Vehicle (UAV) detection algorithms for anti-UAV, this paper proposes a lightweight UAV swarm detection method based on You Only Look Once Version X (YOLOX). This method uses depthwise separable convolution to simplify and optimize the network, and greatly simplifies the total parameters, while the accuracy is only partially reduced. Meanwhile, a Squeeze-and-Extraction (SE) module is introduced into the backbone to improve the model′s ability to extract features; the introduction of a Convolutional Block Attention Module (CBAM) in the feature fusion network makes the network pay more attention to important features and suppress unnecessary features. Furthermore, Distance-IoU (DIoU) is used to replace Intersection over Union (IoU) to calculate the regression loss for model optimization, and data augmentation technology is used to expand the dataset to achieve a better detection effect. The experimental results show that the mean Average Precision (mAP) of the proposed method reaches 82.32%, approximately 2% higher than the baseline model, while the number of parameters is only about 1/10th of that of YOLOX-S, with the size of 3.85 MB. The proposed approach is, thus, a lightweight model with high detection accuracy and suitable for various edge computing devices.

In recent years, with the rapid development of unmanned aerial vehicles (UAV) technology and swarm intelligence technology, hundreds of small-scale and low-cost UAV constitute swarms carry out complex combat tasks in the form of ad hoc networks, which brings great threats and challenges to low-altitude airspace defense. Security requirements for low-altitude airspace defense, using visual detection technology to detect and track incoming UAV swarms, is the premise of anti-UAV strategy. Therefore, this study first collected many UAV swarm videos and manually annotated a dataset named UAVSwarm dataset for UAV swarm detection and tracking; thirteen different scenes and more than nineteen types of UAV were recorded, including 12,598 annotated images—the number of UAV in each sequence is 3 to 23. Then, two advanced depth detection models are used as strong benchmarks, namely Faster R-CNN and YOLOX. Finally, two state-of-the-art multi-object tracking (MOT) models, GNMOT and ByteTrack, are used to conduct comprehensive tests and performance verification on the dataset and evaluation metrics. The experimental results show that the dataset has good availability, consistency, and universality. The UAVSwarm dataset can be widely used in training and testing of various UAV detection tasks and UAV swarm MOT tasks.

In recent years, the increasing number of unmanned aerial vehicles (UAVs) in the low-altitude airspace have not only brought convenience to people’s work and life, but also great threats and challenges. In the process of UAV detection and tracking, there are common problems such as target deformation, target occlusion, and targets being submerged by complex background clutter. This paper proposes an anti-occlusion UAV tracking algorithm for low-altitude complex backgrounds by integrating an attention mechanism that mainly solves the problems of complex backgrounds and occlusion when tracking UAVs. First, extracted features are enhanced by using the SeNet attention mechanism. Second, the occlusion-sensing module is used to judge whether the target is occluded. If the target is not occluded, tracking continues. Otherwise, the LSTM trajectory prediction network is used to predict the UAV position of subsequent frames by using the UAV flight trajectory before occlusion. This study was verified on the OTB-100, GOT-10k and integrated UAV datasets. The accuracy and success rate of integrated UAV datasets were 79% and 50.5% respectively, which were 10.6% and 4.9% higher than those of the SiamCAM algorithm. Experimental results show that the algorithm could robustly track a small UAV in a low-altitude complex background.

Anti unmanned aerial vehicle (UAV) swarm system includes target detection and strike, where detection involves sensing and tracking the target. Aiming at the problems of low detection performance and unstable tracking in traditional methods, a multi-object visual tracking method for low altitude UAV swarm is proposed. Firstly, an Efficient Channel Attention (ECA) module is introduced into the backbone to enhance the ability of feature extraction. Meanwhile, Swin Transformer Block is applied in the neck to enable the model to better capture different local information. Furthermore, alpha Intersection over Union (α-IoU) is used to optimize loss function to facilitate network convergence and improve positioning accuracy. Finally, BYTE strategy is used for data association to increase the integrity of trajectories, thereby achieving accurate and robust tracking of UAV swarm targets. The experiments show that the tracking accuracy on the UAVSwarm test dataset reaches 78.2%, which is about 3% higher than the current advanced model. And it is 12.3%, 6.8%, 10%, and 6% higher than the FairMOT, SORT, OC-SORT, and DeepSORT, proving the effectiveness of the multi-UAV visual tracking method.

The geometric extraction and semantic understanding in bird’s eye view plays an important role in cyberphysical- social systems (CPSS), because it can help human or intelligent agents (IAs) to perceive larger range of environment. Moreover, due to lack of comprehensive dataset from oblique perspective, fog-end deep learning algorithms for this purpose is still in blank. In this paper, we propose a novel method to generate synthetic large-scale dataset for geometric and semantic urban scene understanding from bird’s eye view. There are two main steps involved, one is modeling and the other is rendering, which are processed by CityEngine and UnrealEngine4 respectively. In this way, synthetic aligned multi-model data are obtained efficiently, including spectral images, semantic labels, depth and normal maps. Specifically, terrain elevation, street graph, building style and trees distribution are all randomly generated according realistic situation, a few of handcrafted semantic labels annotated by colors spread throughout the scene, virtual cameras moved according to realistic trajectories of unmanned aerial vehicles (UAVs). For evaluation of practicability of our dataset, we manually labeled tens of aerial images downloaded from internet. And the experiment result show that, in both pure and combined mode, the dataset can improve the performance significantly.

In complex environments such as night, fog and battlefield camouflage, a single camera sensor is not sufficient to reflect scene information, and multi-source sensor can raise environmental awareness. A visible light camera with an infrared sensor is an efficient combination. However, there are huge differences in the inputs from different sensors, and how to fuse the information from two sensors and apply it to a specific task is a problem that needs to be solved. Hence, a multi-source input detection algorithm with the combination of infrared sensors and visible cameras is proposed in this paper. Its purpose is to solve the problem of low detection accuracy of a single sensor in complex and changing environments. First, a differential feature enhancement(DFE) module to enhance features that are constantly degraded during network transmission is designed in this paper. Second, a cross-modal fusion module(CF) to fuse features from multiple sources is designed. Finally, two modules are embedded in a two-stream network. Experiments on the publicly available FLIR and LLVIP datasets show that the algorithm in this paper improves mAP75 by 8.3/4.3 compared to a single-source detector. In some special environments, the module in this paper uses 0.1MB of storage at the cost of a 1.7 mAP boost! Extensive ablation experiments demonstrate that the module proposed in this paper is lightweight, efficient, plug and play.