Multi-agent collaborative perception has shown significant potential in improving environmental awareness, especially in complex and dynamic environments where it enables a more precise and comprehensive understanding. However, existing collaborative perception systems still face several challenges, particularly during feature fusion, where issues such as information redundancy and noise interference arise. To address these challenges, this paper introduces the Dual-Attention Weighting Collaborative Perception (DAW-CoPe) framework. The framework integrates dilated convolutional techniques within the Multi-Scale Spatial Attention Module (MSAM), allowing for the capture of multi-dimensional object features across different scales and the effective integration of richer contextual information, thereby enhancing the model’s perceptual capabilities. Furthermore, the incorporation of local convolutions strengthens the model’s ability to extract fine-grained features, improving its capacity to capture local details during feature fusion and reducing the risk of losing important information. To further enhance perception accuracy, we introduce a dual-attention weighting (DAW) mechanism: a same-level, learnable gating module that jointly reweights channel semantics and multi-scale spatial cues, thereby emphasizing informative channels and critical spatial regions. By suppressing redundant and noisy signals, this mechanism delivers substantial gains in performance and robustness across perception tasks. Experimental results demonstrate that DAW-CoPe outperforms existing state-of-the-art methods on three public datasets, significantly boosting the accuracy, reliability, and adaptability of collaborative perception systems. The code associated with this research will be made publicly available athttps://github.com/HIT-K/DAW-CoPe

With the development of the Internet of Things, cross-view geolocation of satellite images and Unmanned Aerial Vehicles (UAVs) aerial images is gaining attention due to its significant potential in denied-navigation environments under Global Navigation Satellite System (GNSS). However, inadequate feature representation between views, as well as uncertainties in positional offsets and distance scales, constitute key challenges in this field. Existing research primarily focuses on extracting comprehensive and fine-grained feature information. However, effective feature representation and alignment are equally critical. Therefore, a novel visual localization algorithm based on Multi-scale Feature Fusion on improved Transformer architecture (MFFT) is proposed. The algorithm combines Adaptive semantic Feature extraction based on Improved ASPP (AFIA) model and Spatial Pyramid based on Multi-scale Subsampling (SPMS) model, and has strong robustness. MFFT cooperatively uses AFIA and SPMS to successfully achieve an effective balance between feature representation and alignment. Experimental results show that, on the common benchmark data set DenseUAV, compared with the most advanced algorithms in the same field, the R@1 of MFFT algorithm is improved by at least 2.23% and the SDM@1 is improved by at least 1.64%, thus achieving the most advanced cross-view matching performance, verifying that the proposed algorithm has strong competitiveness in UAV visual self-localization tasks, and opening up a new technical path for future UAV autonomous navigation and diversified mission execution in complex environments.

Image retrieval technology has become increasingly important in modern society. Particularly, instance-level image retrieval not only enables fast and accurate image retrieval but also plays a significant role in the field of Visual place recognition (VPR). NetVLAD's (Vector of locally aggregated descriptors) global feature extraction method exhibits excellent performance in the field of image retrieval. Nevertheless, the NetVLAD method is characterized by a long training time, a low recall rate, and a slow method for extracting features. To address this problem, we propose an end-to-end model based on a novel deep neural network for global feature extraction. It exhibits lower computation complexity for high-dimensional features, accelerates feature extraction, and improves the recall rate of image retrieval. We present the following two principal contributions. First, we introduce attention aggregation module based on self-attention mechanism that combines the positional information and confidence scores of local features. And then it enhances local features using the self-attention mechanism. Second, we present the global feature extraction module, At-tnVLAD, based on the principles of the NetVLAD method. It employs a cross-attention mechanism in place of convolution, reducing the number of trainable parameters without affecting the computational complexity. The experimental results indicate that the proposed method can not only speed up training convergence and feature extraction but also improve recall rate.

Nowadays, blind guiding service is highly required by visually impaired people for a normal outdoor life. However, the tactile pavements fail to work for blind guiding when they are covered or blocked, and even become barriers for normal people who are riding bicycles, sliding the trolley cases and so on. Recently, distributed optical fiber vibration sensing (DOFVS) is widely used for electric fence and human activity recognition. In this paper, to our best knowledge, we first propose the optical fiber pavement to replace the tactile pavement for outdoor blind guiding, and provide the positioning method based on phase-sensitive optical time domain reflectometer ($\Phi$-OTDR). By extracting critical gait features from the output intensity signal, we succeed to make accurate classification between the blind pedestrians and the normal ones. Based on signal autocorrelation analysis in both time and space domains, the locations of each target person on the optical fiber pavement are accurately estimated. The experiment results show that our proposed method achieves highly accurate recognition of blind pedestrians and provides their localization estimations accurately.

Aiming to improve the accuracy of visual indoor positioning, an efficient deep multi-patch network based image deblurring algorithm (DMPID) is proposed to eliminate the effect of blurred images on the positioning accuracy. Meanwhile, the self-sorting visual word based image retrieval algorithm and the block based improved eight-point method are also proposed in this paper to improve the retrieval accuracy and positioning accuracy. The proposed image deblurring algorithm adopts the deep multi-patch network and the weight selective sharing scheme to acquire the deblurred images. Then, we propose the self-sorting visual word and add two additional elements representing the relative spatial information into every obtained feature point to utilize the intrinsic relationships between extracted features and physical positions in order to improve the retrieval accuracy. Meanwhile, the visual word filtering is also employed to eliminate redundant visual words and reduce the time consumption of image retrieval. Finally, the position estimation of query camera can be achieved by the epipolar constraint based on the improved eight-point method. Simulation results and performance analysis show that the proposed method can restore the image details effectively and improve the accuracy of visual indoor positioning.

In order to accurately realize the ablation treatment of liver tumors, it is necessary to design a surgical treatment plan for radio frequency ablation (RFA) based on computed tomography (CT) image. Currently, as a popular deep learning method, U-net network is widely used for image segmentation. However, traditional U-net network cannot achieve to segment the liver tumor from the CT images. In order to accomplish the tumor segmentation in CT images and make the design on the RFA treatment, in this paper we propose an improved U-net network. We use two deep convolutional neural networks based on U-net and improve the overall performance of the segmentation network by improving the network model. On the basis of liver tumor segmentation, we finally accomplish the design of RFA for liver tumor removal. Simulation results show that our proposed method can effectively achieve liver tumor segmentation and obtain a good Dice coefficient, thus providing help for the surgical design of RFA for liver tumors.

2D image-based and 3D structure-based are the most two popular vision-based indoor positioning methods. Though 2D method has better efficiency for image retrieval, the number of stored images is usually too many to achieve fast 2D-2D matching, and it cannot obtain high positioning accuracy either. In contrast, 3D method has smaller database and can achieve much higher positioning accuracy with the help of 3D information. However, it needs to construct 3D point cloud map of the positioning environment and spend too much time on 2D-3D matching. Therefore, in order to improve the positioning accuracy and the retrieval efficiency, we propose a vision-based indoor positioning method by joint use 2D images and 3D point cloud map in this paper. In offline stage, we select key images into the offline database based on the 3D point cloud map that is built by RGB-D simultaneous localization and mapping (SLAM), and establish the mapping relation between 2D image local feature points and 3D points. In online stage, we utilize the rough-fine matching strategy to obtain 2D-3D matching pairs between local feature points of the user query image and 3D points in the 3D point cloud map. The user position is finally estimated by the efficient perspective-n-point (EPnP) algorithm. Performance analysis and simulation results indicate that our proposed method can not only achieve high positioning accuracy but also speed up the matching process.

Aiming at the problem of existing bag-of-visual words based image retrieval algorithm, such as poor stability and low retrieval accuracy, a bag-of-visual words based improved image retrieval algorithm (IBVW) is proposed, which extracts features from the images in the database. The approximate K-means algorithm is adopted to cluster the image features into visual words and store them in the database. And then, the feature extraction and the inverted index in the online stage are implemented on the query image to find the images with high similarity in the database. At last, the best matching image is acquired through the similarity calculation, voting scheme and homography based matching algorithm. Simulation results and performance analysis show that the accuracy of our retrieval algorithm is improved by about 10% compared with existing methods for vision indoor positioning.

Quick and accurate floor identification in a multistory building is a challenging task for 3D indoor positioning. The performances of the available methods are low accuracy or even unworkable in the large and complex urban environment due to the noisy received data. Furthermore, the relationship between received data from different reference points is not considered to make the floor identification better. Therefore, in this paper, we focus on improving floor identification accuracy and propose a novel floor identification method. For better describing the property of the signal propagating difference coming from the same base station to different floors, we analyze both the channel characteristics and geographic characteristics and then select five important parameters for floor identification. Based on these parameters, a floor identification method is proposed. We use Denoising Autoencoder (DAE) on these parameters for noise reduction and feature extraction. Then, we use the Long Short-Term Memory (LSTM) on the denoised features for floor identification, which can better explore and utilize data feature relationship. Based on the real cellular network data, the experiment results show that our proposed floor identification method is very accurate for different structural buildings, which outperforms the traditional methods.