Most of the existing infrared and visible image fusion algorithms rely on hand-designed or simple convolution-based fusion strategies. However, these methods cannot explicitly model the contextual relationships between infrared and visible images, thereby limiting their robustness. To this end, we propose a novel Transformer-based feature fusion network for robust image fusion that can explicitly model the contextual relationship between the two modalities. Specifically, our fusion network consists of a detail self-attention module to capture the detail information of each modality and a saliency cross attention module to model contextual relationships between the two modalities. Since these two attention modules can obtain the pixel-level global dependencies, the fusion network has a powerful detail representation ability which is critical to the pixel-level image generation task. Moreover, we propose a deformable convolution-based feature align network to address the slight misaligned problem of the source image pairs, which is beneficial for reducing artifacts. Since there is no ground-truth for the infrared and visible image fusion task, it is essential to train the proposed method in a self-supervised manner. Therefore, we design a self-supervised multi-task loss which contains a structure similarity loss, a frequency consistency loss, and a Fourier spectral consistency loss to train the proposed algorithm. Extensive experimental results on four image fusion benchmarks show that our algorithm obtains competitive performance compared to state-of-the-art algorithms.

An online neutrosophic similarity-based objectness tracking with a weighted multiple instance learning algorithm (NeutWMIL) is proposed. Each training sample is extracted surrounding the object location, and the distribution of these samples is symmetric. To provide a more robust weight for each sample in the positive bag, the asymmetry of the importance of the samples is considered. The neutrosophic similarity-based objectness estimation with object properties (super straddling) is applied. The neutrosophic theory is a new branch of philosophy for dealing with incomplete, indeterminate, and inconsistent information. By considering the surrounding information of the object, a single valued neutrosophic set (SVNS)-based segmentation parameter selection method is proposed, to produce a well-built set of superpixels which can better explain the object area at each frame. Then, the intersection and shape-distance criteria are proposed for weighting each superpixel in the SVNS domain, mainly via three membership functions, T (truth), I (indeterminacy), and F (falsity), for each criterion. After filtering out the superpixels with low response, the newly defined neutrosophic weights are utilized for weighting each sample. Furthermore, the objectness estimation information is also applied for estimating and alleviating the problem of tracking drift. Experimental results on challenging benchmark video sequences reveal the superior performance of our algorithm when confronting appearance changes and background clutters.

Visual-inertial SLAM systems achieve highly accurate estimation of camera motion and 3D representation of the environment. Most of the existing methods rely on points by feature matching or direct image alignment using photo-consistency constraints. The performance of these methods usually decreases when facing low textured environments. In addition, lines are also very common in man-made environments and provide geometrical structure information of the environment. In this paper, we increase the robustness of visual-inertial SLAM system to handle these situations by using both points and lines. Our method, implemented based on ORB-SLAM2, makes the combination of points, lines and IMU measurements in an effective way by selecting keyframes very carefully and handling the outlier lines efficiently. The cost function of bundle adjustment is formed by point, line reprojection errors and IMU residual errors. We derive the Jacobian matrices of line reprojection errors with respect to the 3D endpoints of line segments and camera motion. Loop closure detection is decided by both point and line features using the bag-of-words approach. Our method is evaluated on the public EuRoc dataset, and compared with the state-of-the-art visual-inertial fusion methods. Experimental results show that our method achieves the highest accuracy on most of testing sequences, especially in some challengeable situations such as low textured and illumination changing environments.

The acceleration of urbanization has led to increasingly severe traffic congestion, creating an urgent need for effective traffic signal control strategies to improve road efficiency. This paper proposes an adaptive traffic signal control method based on offline reinforcement learning (Offline RL) to address the limitations of traditional fixed-time signal control methods. By monitoring key parameters such as real-time traffic flow and queue length, the proposed method dynamically adjusts signal phases and durations in response to rapidly changing traffic conditions. At the core of this research is the design of a model named SD3-Light, which leverages advanced offline reinforcement learning to predict the optimal signal phase sequences and their durations based on real-time intersection state features. Additionally, this paper constructs a comprehensive offline dataset, which enables the model to be trained without relying on real-time traffic data, thereby reducing costs and improving the model’s generalization ability. Experiments conducted on real-world traffic datasets demonstrate the effectiveness of the proposed method in reducing the average travel time. Comparisons with several existing methods highlight the clear advantages of our approach in enhancing traffic management efficiency.

Spatio-temporal prediction is crucial in intelligent transportation systems (ITS) to enhance operational efficiency and safety. Although Transformer-based models have significantly advanced spatio-temporal prediction performance, recent research underscores the importance of learning dynamic spatio-temporal dependencies for these tasks. This paper introduces the Dual-Gated Graph Convolutional Recurrent Unit with Integrated Graph Learning (DG3L), a framework specifically designed to address the complex demands of spatio-temporal prediction. The DG3L model includes a memory-based graph learning module capable of generating dynamic graphs to accurately reflect ongoing changes in spatio-temporal dependencies. By integrating the strengths of Transformer and Graph Convolutional Recurrent Unit (GCRU) technologies within its Dual-Gated Graph Convolutional Recurrent Unit architecture, DG3L provides a mechanism for fusing Transformer features with contextual features from recurrent units. In practical applications, DG3L acts as an advanced representation learning module, delivering highly accurate context features for complex downstream tasks in ITS.

Transformer-based methods have recently demonstrated their potential in time series forecasting problems. However, the mainstream approach, primarily utilizing attention to model inter-step correlation in the time domain, is constrained by two significant issues that lead to ineffective and inefficient multivariate forecasting. The first is that key representations in the time domain are scattered and sparse, resulting in parameter bloat and increased difficulty in capturing time dependencies. The second is that treating time step points as uniformly embedded tokens leads to the erasure of inter-variate correlations. To address these challenges, we propose a frequency-wise and variables-oriented transformer-based method. This method leverages the intrinsic conjugate symmetry in the frequency domain, enabling compact frequency domain representations that naturally mix information across time points while reducing spatio-temporal costs. Multivariate inter-correlations can also be captured from similar frequency domain components, which enhances the variables-oriented attention mechanism modeling capability. Further, we employ both polar and complex domain perspectives to enrich the frequency domain representations and decode complicated temporal patterns. We propose frequency-enhanced independent representation multi-head attention (FIR-Attention) to leverage these advantages for improved multivariate interaction. Techniques such as cutting-off frequency and equivalent mapping are used to ensure the model’s lightweight nature. Extensive experiments on eight mainstream datasets show that our approach achieves first-rate satisfactory results and, importantly, requires only one percent of the spatio-temporal cost of mainstream methods.