Moving target tracking is one of the most active applications in the field of unmanned aerial vehicles (UAVs). Due to their high adaptability, UAVs equipped with multisensor vision tracking techniques can perform various tasks in hazardous environments. However, varying application requirements and scenarios result in different tracking focuses. This article reviews the current research status of UAV moving target tracking with tracking requirements in mind. First, this article proposes a taxonomy that categorizes target tracking techniques into three categories: detection tracking, following tracking, and cooperative tracking, which based on the varying requirements of target tracking in different scenarios. Furthermore, the current UAV algorithms for tracking moving targets are classified and summarized based on different application scenarios. Finally, this article conducts research and analysis on experimental results, funded projects, and competitions to compare the real-world application effects and current progress of UAV tracking methods in relevant fields. In addition, it discusses the challenges and future research directions for UAVs in tracking.

Simultaneous localization and mapping (SLAM) has a wide range for applications in mobile robotics. Lightweight and inexpensive vision sensors have been widely used for localization in GPS-denied or weak GPS environments. Mobile robots not only estimate their pose, but also correct their position according to the environment, so a proper mathematical model is required to obtain the state of robots in their circumstances. Usually, filter-based SLAM/VO regards the model as a Gaussian distribution in the mapping thread, which deals with the complicated relationship between mean and covariance. The covariance in SLAM or VO represents the uncertainty of map points. Therefore, the methods, such as probability theory and information theory play a significant role in estimating the uncertainty. In this paper, we combine information theory with classical visual odometry (SVO) and take Jensen-Shannon divergence (JS divergence) instead of Kullback-Leibler divergence (KL divergence) to estimate the uncertainty of depth. A more suitable methodology for SVO is that explores to improve the accuracy and robustness of mobile devices in unknown environments. Meanwhile, this paper aims to efficiently utilize small portability for location and provide a priori knowledge of the latter application scenario. Therefore, combined with SVO, JS divergence is implemented, which has been realized. It not only has the property of accurate distinction of outliers, but also converges the inliers quickly. Simultaneously, the results show, under the same computational simulation, that SVO combined with JS divergence can more accurately locate its state in the environment than the combination with KL divergence.

In this study, we introduce an Unmanned Aerial Vehicle (UAV) centric visual semantic simultaneous localization and mapping (SLAM) framework that integrates RGB–D cameras, inertial measurement units (IMUs), and a 5G–enabled remote interaction module. Our system addresses three critical limitations in existing approaches: (1) Distance constraints in remote operations; (2) Static map assumptions in dynamic environments; and (3) High–dimensional perception requirements for UAV–based applications. By combining YOLO–based object detection with epipolar–constraint-based dynamic feature removal, our method achieves real-time semantic mapping while rejecting motion artifacts. The framework further incorporates a dual–channel communication architecture to enable seamless human–in–the–loop control over UAV–Unmanned Ground Vehicle (UGV) teams in large–scale scenarios. Experimental validation across indoor and outdoor environments indicates that the system can achieve a detection rate of up to 75 frames per second (FPS) on an NVIDIA Jetson AGX Xavier using YOLO–FASTEST, ensuring the rapid identification of dynamic objects. In dynamic scenarios, the localization accuracy attains an average absolute pose error (APE) of 0.1275 m. This outperforms state–of–the–art methods like Dynamic–VINS (0.211 m) and ORB–SLAM3 (0.148 m) on the EuRoC MAV Dataset. The dual-channel communication architecture (Web Real–Time Communication (WebRTC) for video and Message Queuing Telemetry Transport (MQTT) for telemetry) reduces bandwidth consumption by 65% compared to traditional TCP–based protocols. Moreover, our hybrid dynamic feature filtering can reject 89% of dynamic features in occluded scenarios, guaranteeing accurate mapping in complex environments. Our framework represents a significant advancement in enabling intelligent UAVs/UGVs to navigate and interact in complex, dynamic environments, offering real-time semantic understanding and accurate localization.

Designing complex UAV swarm performances like drone light shows and aerial displays demands both geometric precision and expressive flexibility, yet existing approaches rely on manually crafted trajectories, requiring substantial manual effort and limiting creative flexibility. While recent large language models (LLMs) offer an opportunity to bridge natural language intent and trajectory generation, general-purpose LLMs lack the geometric reasoning needed to produce spatially consistent paths. We propose an interactive end-to-end framework that translates natural language and speech instructions into executable multi-agent flight trajectories, in which language-specified 2D patterns are mapped to 3D flight coordinates at assigned altitudes. Our approach introduces SVG path commands as a vectorized representation that is both LLM-friendly and geometrically precise, serving as a semantic–geometric bridge. To equip LLMs with geometric awareness, we construct a difficulty-tiered dataset and perform supervised fine-tuning (SFT) with LoRA, complemented by greedy nearest-neighbor target assignment and BallTree-accelerated artificial potential field (APF) deconfliction for scalable execution. Experiments across ten frontier LLMs reveal a pronounced capability cliff: the best Tier 1 accuracy ranges from 88% to 92% across the two judges, whereas Tier 2 accuracies range from 0% to 26% across models and judges. After domain enhancement, our SFT-enhanced 4B-parameter model with few-shot prompting achieves 80% on basic shapes and 22% on text outlines, giving 51% overall accuracy versus 43% for the 9B model under the tested configurations. Different LoRA ranks and training token cutoffs confound this scale comparison. The Llama comparison also shows that fine-tuning gains do not generalize across the tested model families and settings. The 2 m minimum-separation threshold is satisfied in 95–100% of randomized-seed runs for swarms up to $N=50$ under the tested grid-launch protocol and in all tested phased-deployment runs for the $N=100$ dense text formation; these are empirical results rather than formal safety guarantees. The structural validity of the recommended fine-tuned configuration reaches 100% on both tiers (versus 0–34% for the two Qwen pretrained baseline configurations), and AirSim physics-based simulation validation reports a mean formation error below 1 m for five of six tested patterns, averaged from the minimum-error frame to the end of each run, with one transient separation violation observed in a single pattern. For small and moderate swarms, the offline design cycle takes approximately 30–74 s and is dominated by one LLM inference call; the densest $N=200$ case additionally requires up to 48.7 s for APF convergence.