While Transformer-based visual trackers excel in spatial feature representation, the quadratic computational complexity of the self-attention mechanism limits their efficiency in modeling long sequences. Furthermore, in long-term tracking scenarios, conventional single-memory update strategies often suffer from implicit cumulative errors or explicit appearance memory contamination from background noise, making it difficult for the model to robustly recover from severe occlusions. To address these limitations, this paper proposes HMTrack, a visual object tracking method based on a hybrid spatio-temporal backbone network and dual memory mechanisms. First, we construct a spatio-temporal hybrid backbone network, M-iTPN, which interleaves Mamba modules into the hierarchical structure of ViT. This design maintains fine-grained spatial attention while introducing efficient temporal aggregation at selected stages through a selective scanning mechanism. Second, to adaptively accommodate the short-term motion and long-term appearance evolution of the target, we propose a dual-memory mechanism: on the one hand, temporal tokens are used as an implicit temporal-state memory to inject historical context and maintain trajectory smoothness; on the other hand, an adaptive tri-state memory bank is used as an explicit reliability-aware appearance memory to provide identity anchoring, short-term adaptation, and high-confidence recovery references, thereby reducing memory drift. Extensive experiments on seven mainstream benchmark datasets demonstrate that HMTrack achieves an excellent balance between accuracy and speed. Particularly on the LaSOT dataset, our method achieves an AUC score of 72.3% and an inference speed of 40 FPS, demonstrating highly competitive performance compared to recent advanced methods. Our source code and pre-trained models are available at: https://github.com/djy-cv/HMTrack.

Cooperative localization is critical for UAV swarm operations in GNSS-denied environments. The backbone-listener scheme, using a small subset of agents as active backbone nodes and others as passive listeners, offers notable advantages in reducing communication overhead and enhancing swarm scalability. Building on this scheme, we propose a formation-constrained cooperative localization method to improve accuracy by integrating known formation geometry into the localization process. First, backbone node selection uses a formation-constrained greedy node activation (GNA) strategy with weighted distance fusion, combining measured and ideal formation distances to enable near-optimal selection aligned with formation structure. Second, listener node localization incorporates formation constraints into Chan’s algorithm, paired with angle-of-arrival (AOA) refinement, to ensure estimated positions match expected inter-agent distances. Third, global optimization uses a gradient descent-based refinement to enforce formation constraints across all agent positions. Our theoretical derivations and simulations are limited to the two-dimensional (2D) case. Simulation results validate the proposed method’s improved success rate, reliability, and stability. Its effectiveness is demonstrated across various formation types, with robust adaptability to asymmetric geometries shown to be a valuable feature for practical deployment.

To tackle the frequent missed and false detection issues arising from the tiny scale of objects and strong background clutter in UAV aerial photography scenarios, this paper proposes a novel algorithm named SODet-YOLO for UAV aerial imagery. First, to effectively extract the features of aerial objects and alleviate background interference, we integrate a high-resolution detection head denoted as P2 into the YOLO11n, which is connected to the feature layer from the second downsampling stage of the Backbone and Neck networks, we design the Fine-Grained Aggregation-Asymptotic Feature Pyramid Network (FGA-AFPN) to realize adequate fusion of feature information at different levels. Second, we redesign the original C3k2 module by embedding the Inception Depthwise Convolution (IDC). This design effectively expands the receptive field, enriches multi-scale contextual feature extraction, and mitigates adverse interference from complex background clutter. In addition, a novel IoU loss function named MPDInterpIoU is proposed by combining InterpIoU with MPDIoU. This function promotes faster convergence at the early learning stage and optimizes detection-related performance. Finally, the Parallelized Patch-aware Attention (PPA) is incorporated before the downsampling module to preserve the key features of small objects throughout multiple downsampling steps. The experimental findings validate that SODet-YOLO achieves an mAP@0.5 score of 41.487% on the VisDrone2019 object detection dataset, representing an 8.92% performance enhancement relative to the baseline YOLO11n model. However, the computational cost increases moderately, with the number of parameters increasing by 1.08 M, the computational complexity increasing by 26.1 GFLOPs, and the average inference time growing by 34.7 ms.

In response to the difficulty of detecting small-sized targets in drone aerial scenes, an improved YOLOv11n small target detection algorithm based on progressive feature pyramid network is proposed. Firstly, a detection head is added to the high-resolution feature layer of the backbone network, and feature information between non adjacent layers is fused through an asymptotic feature pyramid network (AFPN) to alleviate the problem of feature information loss caused by downsampling and reduce conflicts between cross level feature information. Secondly, an improved SPPF based on spatial channel attention mechanism is used to replace the original SPPF, highlighting important features and suppressing irrelevant feature information through spatial and channel attention mechanisms, further enhancing the model's performance. Finally, combining MPDIoU and InnerIoU improves detection accuracy. The experiment shows that the improved algorithm achieves a performance of 38.537 on the Visdrone2019 dataset mAP@0.5 Compared to the benchmark model, it has increased by 5.97%.

The video object segmentation algorithm based on memory networks stores the information of the target object through the maintained external memory inventory. As the segmentation progresses, the size of the memory inventory will continue to increase, leading to redundancy of feature information and affecting the execution efficiency of the algorithm. In addition, the key value pairs stored in the memory library are subjected to channel dimension reduction using standard convolution, resulting in insufficient representation ability of target object features. In response to the above issues, this chapter proposes a video object segmentation algorithm based on feature compression and attention correction, constructing a reliable and effective memory library to ensure efficient storage and updating of target object information, thereby reducing computational complexity and storage consumption. A dual attention mechanism based on spatial and channel dimensions was proposed to correct feature information and enhance the representation ability of features. A large number of experiments have shown that the proposed algorithm demonstrates reliable competitiveness compared to other mainstream algorithms in recent years.

In order to solve the problems of low detection accuracy and poor real-time performance caused by small size, dense distribution and complex background of UAV aerial vehicle detection in UAV aerial vehicle detection scenes, this paper proposes an improved UAV aerial vehicle detection algorithm based on YOLOv11n. Firstly, a detection head was added to the high-resolution feature layer of the YOlOv11n backbone network to reduce the problem of small target information loss caused by the reduction of resolution after downsampling, and improve the detection accuracy of small target. Secondly, the Inner-IoU Loss is used to replace the traditional IoU Loss, and the auxiliary box is used to accelerate the convergence of the regression process. In addition, the Multi-scale Attention Aggregation Module (MSAA) is introduced to fuse multi-scale feature information by using spatial and channel attention mechanisms, which can improve the effect of multi-scale spatial and channel fusion while reducing background interference. Experiments show that the improved algorithm achieves a of 38.302

The typical shortcoming of STM (Space-Time Memory Network) mode video object segmentation algorithms is their high segmentation performance coupled with slow processing speeds, which poses challenges in meeting real-world application demands. In this work, we propose using an online knowledge distillation method to develop a lightweight video segmentation algorithm based on the STM mode, achieving fast segmentation while maintaining performance. Specifically, we utilize a novel adaptive learning rate to tackle the issue of inverse learning during distillation. Subsequently, we introduce a Smooth Block mechanism to reduce the impact of structural disparities between the teacher and student models on distillation outcomes. Moreover, to reduce the fitting difficulty of the student model on single-frame features, we design the Space-Time Feature Fusion (STFF) module to provide appearance and position priors for the feature fitting process of each frame. Finally, we employ a simple Discriminator module for adversarial training with the student model, to encourage the student model to learn the feature distribution of the teacher model. Extensive experiments show that our algorithm attains performance comparable to the current state-of-the-art on both DAVIS and YouTube datasets, despite running up to x 4 faster, with x 20 fewer parameters and x 30 fewer GFLOPS.

Visual object tracking (VOT) algorithms based on Transformer have excellent tracking performance, but they often neglect historical image information and are difficult to avoid interference from similar objects. To address these issues, a Visual object tracking algorithm based on Dynamic Template and Position Enhancement is proposed in this paper. On the one hand, Dynamic Template (DT) is added to the algorithm to resolve the problem of insufficient representation of target appearance information by a single template, to obtain high-quality dynamic templates, an update strategy is proposed that is based on similarity determination at a fixed number of frames. On the other hand, a Positional Enhancement (PE) module is added to the Transformer's self-attention computation process to reduce the interference of background information on foreground information. Extensive experiments were conducted on several datasets to evaluate the proposed algorithm, which demonstrated excellent performance. The proposed algorithm also achieves an inference speed of 43 FPS. The experimental results show that the proposed algorithm can effectively handle tracking task in complex scenes.

To address the problem of target disappearance and reappearance in the filed of view frequently, a re-detection method is required to search for the target in the entire image. However, this search strategy will bring more interference from similar objects, and also makes it harder to detect small objects. Therefore, we designed a deep learning model, which is a global re-detection method based on multi-scale feature fusion and attention mechanisms. Firstly, the proposed method uses ResNet50 to extract shallow and deep features of the template and search region. Then, a multi-scale feature fusion module is used to fuse shallow and deep features, which enhance the accurate localization ability for small objects. Finally, A target-aware proposal enhancement module is proposed to enhance the response of proposal, especially the parts related with template, which filters out more accurate proposals and improves the ability of re-detection methods to handle interference from similar objects. The proposed method in this paper is carried on several datasets such as UAV20L, OTB, etc., and the experimental results show that the tracking performance of the proposed method is significantly improved.

Recently, deep convolutional neural networks (DCNN) have been widely used in semantic segmentation tasks and have achieved high segmentation accuracy. However, most algorithms based on DCNN have high computational complexity, making them unsuitable for real‐time segmentation. To solve this problem, this paper proposes a real‐time semantic segmentation algorithm based on the STDC network. The algorithm adopts an “encoder–decoder” embedded in a U‐shaped architecture to realize real‐time segmentation while maintaining high accuracy. Following the encoder, a mixed pooling attention module is designed to expand the receptive field, enhancing the network model’s learning ability in complex scenarios. Then, a feature fusion module is used for combining features from different stages, and channel attention based on atrous convolution is employed to expand the receptive field and avoid dimensionality reduction learning. Finally, a Tversky‐based detail loss function is used to encode more spatial details. The proposed algorithm was extensively tested on the challenging Cityscapes and CamVid datasets, and the experimental results showed that the proposed algorithm obtained 76.4% and 72.8% of mIoU, respectively. Meanwhile, our algorithm achieves 105.2 FPS and 165.6 FPS inference speed with a single NVIDIA GTX 1080Ti GPU, meeting the real‐time segmentation requirements. The proposed algorithm can conduct real‐time segmentation while maintaining high accuracy, achieving a good balance between accuracy and speed.

Verifying whether the tracking result drifts during long-term tracking is a critical challenge. It’s difficult to select the optimal threshold in traditional threshold-based object drift verification criteria, and existing threshold-free object drift verification networks perform poorly in complex scenarios. To address these issues, we propose an object drift verification network based on multi-scale feature fusion and dual-template, using static and dynamic templates for joint verification. During the feature extraction stage, a multi-scale feature fusion module is introduced to adapt to changes in the object’s scale. Additionally, A template update strategy is devised to obtain high-quality dynamic templates for effective object drift verification. This network doesn’t require manual threshold setting and can be used as a plug-and-play module combined with a short-term visual tracking algorithm and a global re-detection module for long-term tracking. The proposed network forms four long-term tracking algorithms by integrating with four short-term visual tracking algorithms (DiMP50, PrDiMP, TrDiMP, and TransT). Extensive experiments on datasets like LaSOT, UAV20L, VOT2018-LT, and VOT2020-LT demonstrate significant improvements in long-term tracking performance. On the UAV20L dataset, the success rate and precision improved by 8.5

Multi-object tracking (MOT) detects multiple targets in an image and assigns a unique identifier to each target. However, challenges such as rapid motion, occlusion, and camera motion in the tracking scene may lead to identity switches (IDs) and missing trajectory problems, which degrade the performance of the tracker. To address these issues, this paper presents an MOT algorithm based on an interactive attention network and adaptive trajectory reconnection. First, an interactive attention network was created to learn the features for two different tasks of detection and tracking to alleviate feature conflicts in order to extract sufficient feature information. A new cost matrix was then designed to fuse the motion and feature information, thereby reducing the number of IDs. Meanwhile, the extreme gradient boosting reconnection module was used to achieve adaptive trajectory reconnection and reduce missing trajectories. The proposed algorithm achieved 61.5% and 55.4% HOTA using the standard MOT17 and MOT20 datasets, respectively. In comparison to FairMOT, our algorithm showcased notable enhancements of 3% and 1.6% on these datasets. Furthermore, when compared to state-of-the-art algorithms, the proposed algorithm demonstrated superior tracking performance.

The Transformer-based visual tracking has demonstrated exceptional performance, but there is still space for further improvement in target feature expression. To address this issue, this paper proposes a Sparse Transformer visual tracking network based on second-order attention to enhance the feature expression capability of the tracking algorithm. Firstly, the spatio-temporal motion information is integrated in this paper to model the motion of the target in the video sequence, thereby enhancing the feature expression of the target. Secondly, the proposed method improves the Transformer structure in the visual tracking by utilizing a sparse self-attention mechanism that focuses on the most crucial information in the target area. More importantly, to further enhance the discriminative ability of the tracker for the target, a mixed second-order pooling (MSOP) module is embedded in the encoder-decoder structure. The proposed method has been extensively evaluated on multiple datasets, achieving success rates of 70.9%, 67.9%, and 64.8% on OTB100, UAV123, and LaSOT, respectively. Furthermore, the method obtains EAO score of 0.485 on VOT2018. The experimental results demonstrate that the proposed method has better tracking performance and is more robust in various complex scenarios.

The discriminative correlation filter (DCF) is commonly used in aerial object tracking due to its high tracking accuracy and computing speed. However, when similar object disturbances emerge in the background, the response map will generate sub-peaks, which may eventually lead to tracking failure. Meanwhile, the lack of attention to the tracked object can also cause tracking performance degradation. To these concerns, this paper proposes a novel correlation filter algorithm for real-time aerial tracking based on spatial disturbance suppression and object saliency-aware, i.e., SOCF. Firstly, this paper designs a novel spatial disturbance suppression strategy. Using the temporal information in the historical response maps, we construct a context response map, deviating it from the current response map to detect disturbance information in the background. Then, construct a spatial interference map, divide it into n×n non-overlapping regions, and suppress the negative samples in the disturbance region within the main regression. Furthermore, an object saliency-aware strategy is proposed, using a saliency detection algorithm to calculate the object-aware mask and multiplying it with the detection filter to obtain the object-aware filter. By constructing the object-aware regularization in the training phase, the trained detection filter focuses more on the object itself and can effectively separate the object from the background. Extensive experiments on four widely used unmanned aerial vehicle (UAV) datasets demonstrate that the proposed SOCF tracker achieves high tracking performance. Meanwhile, our tracker can maintain real-time aerial tracking at 48 FPS on a single CPU.

This paper investigates a multi-objective job shop scheduling problem with manual loading and unloading tasks (MOJSSPLU) aiming to minimize makespan and total workload simultaneously. Taking human factors into consideration achieves a better balance between production efficiency and resource utilization. To address this problem, we first establish a bi-objective mixed-integer linear programming (MILP) model. Afterward, we combine the artificial bee colony algorithm with the grid technique to develop a grid-based artificial bee colony algorithm (GABC) for solving MOJSSPLU. The algorithm uses a decomposition approach with the earliest-shortest dispatching rule to reduce the complexity of MOJSSPLU. A grid coordinate system is constructed to divide the objective space into cells, facilitating individuals’ location and evaluation in the decision space. Moreover, we focus on improving the cooperation between the employed and onlooker bees and design a mathematical formula based on priority weights to generate onlooker bees to maximize information utilization and improve the exploration and exploitation capabilities of the algorithm. To ensure that the most promising individuals are retained and protected from being eliminated during the evolutionary, we propose an elitist non-dominated solutions reservation strategy for improving the diversity of the population. We further propose a dynamic adjustment of the grid-division method that focuses on improving the exploitation capacity of the algorithm in the objective space to enhance the diversity and convergence of the population. To test the performance of the GABC algorithm, we compare it to the MILP approach and other algorithms using modified benchmark instances. Furthermore, we conduct the Friedman and Wilcoxon nonparametric tests to validate the comparison results among the different algorithms. The computational results demonstrate that the algorithm GABC is efficient.

In modern warfare, the ability to make timely and informed decisions is critical for mission success. The traditional OODA (Observe, Orient, Decide, Act) loop provides a framework for combating process analysis, but in complex and dynamic environments, collaboration among diverse stakeholders becomes essential. This paper proposes a collaborative OODA (Colla-OODA) model for military combating process analysis. The Colla-OODA model integrates collaboration into each phase of the decision-making process, emphasizing the importance of information sharing, collective analysis, and synchronized action among military units, branches, and allied forces. Sensor fusion plays a central role in enhancing situational awareness by integrating data from various sources such as satellites, drones, ground sensors, and intelligence reports. Through advanced data processing and fusion algorithms, the Colla-OODA model aims to provide decision-makers with a holistic view of the operational environment, enabling more accurate assessments and rapid response to emerging threats and opportunities. The Colla-OODA model is evaluated through comprehensive modeling and simulation analyses, considering factors such as decision-making speed, accuracy, and agility in dynamic scenarios. The findings of this study will contribute to the development of innovative approaches to military decision-making, with potential applications in future battlefield operations and strategic planning efforts.

In long-term visual tracking, most of the target loss discriminative methods require artificially determined thresholds, and the selection of optimal thresholds is usually difficult, resulting in weak generalization ability of long-term tracking algorithms. A target drift Discriminative Network (DNet) that does not require artificially selected thresholds is proposed. The network adopts Siamese structure and uses both static and dynamic templates to determine whether the tracking results are lost or not. Among them, the introduction of dynamic templates effectively improves the algorithm’s ability to adapt to changes in target appearance. In order to train the proposed target drift discriminative network, a sample-rich dataset is established. To verify the effectiveness of the proposed network, a complete long-term tracking algorithm is constructed in this paper by combining this network with the base tracker and the re-detection module. It is tested on classical visual tracking datasets such as UAV20L, LaSOT, VOT2018-LT and VOT2020-LT. The experimental results show that compared with the base tracker, the tracking accuracy and success rate are improved by 10.4% and 7.5% on UAV20L dataset, respectively.

In long-term tracking datasets, it is often encountered with complex scenarios where the object is occluded or temporarily leaves the field of view and then reappears. To cope with these challenges, short-term visual tracking algorithms are usually combined with re-detection methods to construct long-term visual tracking algorithms. However, the existing re-detection methods have limitations in tracking accuracy, which restricts the performance of long-term visual tracking algorithms to a certain extent. To address this problem, a global re-detection method based on feature interaction Siamese network is proposed. Firstly, the ResNet50 is used to extract multi-layer features from template images and search images. Secondly, by introducing the cross-scale feature interaction module, the deep feature interaction unit, and the shallow feature interaction unit, a richer and more comprehensive feature representation is obtained. Finally, the short-term tracking algorithm DiMP is combined with the proposed global re-detection method to form a long-term visual tracking algorithm. The long-term visual tracking algorithm is tested on several publicly available datasets, including UAV20L, LaSOT, UAV123 and VOT2020-LT. The experimental results show that the addition of the global re-detection method improves the accuracy and robustness of the long-term tracking algorithm.

Aiming at the limitations of visible images in object detection, this paper proposes a dual-branch network object detection algorithm based on dual-modality fusion of visible and infrared images. Based on YOLOv7-s, the algorithm firstly introduces a spatial attention module to enhance the model’s ability of capturing key features; secondly, to resolve the problem of inconsistent object sizes, a visible multi-scale feature fusion module is proposed, meanwhile, the structure of the SimCSPSPPF module (an improved spatial pyramid pooling module) from YOLOv6 is adopted to construct an infrared multi-scale feature fusion module to efficiently extract multi-scale features from infrared images; finally, a cross-modal feature fusion module is proposed to fuse corresponding scale features from visible and infrared images. The proposed algorithm is tested on KAIST, FLIR, and GIR datasets, experimental results show that the proposed algorithm has better performance, compared with the YOLOv7-s algorithm to detect visible and infrared images separately on the KAIST dataset, the detection accuracy is improved by 18.0 and 5.1

This paper investigates an energy-efficient scheduling problem on unrelated parallel machines considering general position-based deterioration which arises from the labour-intensive textile industry. The actual processing time of a job is not only associated with the job and the machine but also with its position in the processing sequence. The objective is to minimise the total energy consumption with a bounded makespan. To address this problem, we first establish a mixed-integer linear programming (MILP) model. Afterwards, the initial model is improved by deriving lower and upper bounds on the makespan, and an upper bound on the number of jobs processed on each machine. We also develop an iterative heuristic embedded with a variable neighbourhood search procedure (IHVNS). The algorithm obtains initial solutions iteratively by solving assignment problems and then repairs and improves them with the VNS procedure. Computational results demonstrate that the improved model is up to 230 times faster than the original one. Moreover, the proposed heuristic yields excellent solutions with average gaps of less than 0.73% for large-scale instances. Especially, the results reveal that the IHVNS algorithm is more suitable than MILP models for solving large-scale problems with tight makespan restrictions.