3D multi-object tracking has been extensively applied in areas such as autonomous driving, uncrewed aerial vehicles, and robots. However, existing 3D multi-object tracking methods still face challenges including inaccurately fitting the true motion of objects, insufficient utilization of trajectory information, and trajectory drift. To address these issues, we propose a 3D multi-object tracking framework driven by multi-level association and intelligent filtering. We design an adaptive prediction module for state estimation that reduces noise and error in prediction information through estimation and smoothing processes, thereby enhancing the stability and accuracy of state prediction. Subsequently, a multi-level trajectory integration-guided association strategy is introduced. This strategy integrates detection results, trajectory data, and potential real-world states of the objects. It minimizes incorrect associations and identity switches, thus achieving more accurate and robust data association. Finally, we propose a quality-aware intelligent filtering module for trajectory correction. This module assesses the quality of all matched detection-trajectory pairs and applies regression correction to low-quality drift detections, effectively reducing trajectory fragmentization and identity switches. On the KITTI test dataset, our method achieves 80.64% HOTA and 53.68% HOTA for the car and pedestrian categories, respectively, while reducing identity switches to 50 and 97. These results demonstrate that, compared with existing methods, our method delivers superior tracking performance and exhibits stronger robustness in handling challenges such as large inter-frame displacements, long-term occlusions, and identity switches.

In autonomous driving, accurate 3-D multiobject tracking (MOT) plays a key role in ensuring vehicle safety. However, due to the complexity of the environment, existing methods still face many challenges when dealing with long-distance objects, partial occlusions, and interference from similar categories. To tackle these challenges, we propose a 3-D MOT framework based on a voxel masking encoder (VME) and a deep hashing paradigm (DHP). We introduce a masking strategy that processes voxel features from near to far while maintaining feature sparsity, effectively capturing global contextual information between spatial features. Simultaneously, DHP is utilized to generate image hash codes and compute their hamming distance from the category hash codes. This process effectively distinguishes between object categories and thus avoids cross-category object dissociation. In addition, we propose a distance optimization matching (DOM) method that uses geometric dimensions and spatial distances to build a cost matrix, achieving more efficient and precise object associations. Results from experiments conducted on the KITTI dataset reveal that our framework delivers outstanding tracking performance, surpassing other advanced methods in tracking accuracy. The code is released at https://github.com/lsy-collab/VD-MOT.

Existing Large Multimodal Models (LMMs) demonstrate excellent performance in handling visual tasks in everyday scenarios. However, they still face challenges in understanding structured images, such as flowcharts and organizational charts, which are characterized by text-rich and complex hierarchical components. In this paper, we propose SiQA, a knowledge construction and Retrieval-Augmented Generation(RAG)-based multimodal Question-Answering model designed for Structured Images. SiQA operates in three stages: Knowledge Graph (KG) generation, retrieval-augmented, and answer generation. First, a KG representing the semantics of the structured images is generated through component analysis. We then performed similarity retrieval between the KG and queries, using a node-first algorithm to construct the most relevant subgraph. Finally, after performing an encoding alignment on the multimodal information, it is fed into the LLM to generate the answer. Additionally, we introduce a new dataset, OCQA1, which includes 5,112 questions derived from 1,000 Organizational Charts. We evaluated SiQA’s structured image detection and question-answering capabilities on the FD-DETR (a flowchart dataset) and SCQA, and verified its effectiveness and strong generalization ability through comparisons with existing state-of-the-art (SOTA) methods.

3D multi-object tracking is an important component of autonomous driving technology. Recent 3D multi-object tracking methods still suffer from issues such as information loss during the fusion of multimodal features, weak discriminative power of the association matrix, and poor robustness of single similarity measure. To address these problems, this paper proposes a Multimodal Collaboration-guided Cascade Association network for 3D multi-object tracking (MCCA-MOT). We design a point cloud feature adaptive diffusion fusion module. This module utilizes inverse distance weighting aggregation diffusion technology to address the issue of information loss during the feature fusion process. This enhances the tracking performance of small objects. Secondly, we propose a dynamic sampling feature cooperative fusion module. This module performs fine-grained local-global feature cooperative fusion based on dynamic sampling, enhancing the distinctiveness of object features. It improves the tracking capability of occluded objects. Finally, in the multi-similarity measure-driven cascading association module, we construct a more discriminative association matrix using multiple types of information and design a cascading strategy. This strategy applies different similarity measures at different association stages for objects with ambiguous features. This reduces identity switches and trajectory fragmentation. Extensive experiments on the KITTI dataset demonstrate the superiority of our method in various performance metrics. Our detailed implementations can be obtained athttps://github.com/yuanfuture/MCCA-MOT.

The rapid expansion of Internet of Vehicles (IoV) networks has facilitated high throughput and reliable vehicular communications. Mobile vehicular networks face the challenges: diversification of network equipment, user mobility, and the broadcast nature of wireless channels, so physical layer security modeling of IoV communication systems has become important. The complexity of wireless communication channels makes real-time prediction of secrecy performance challenging. This paper presents an analysis of secrecy performance for mobile vehicular networks. To ensure data secure transmission, we have employed the decode-and-forward (DF) relaying scheme. The signal-to-noise ratio (SNR) of the effective end-to-end link is employed to obtain the mathematical expression results, which can evaluate the secrecy performance. The theoretical secrecy performance is confirmed via simulation. Then, we design a dense-inception convolution neural network (DI-CNN) model, and propose a DI-CNN-based intelligent prediction algorithm.Transformer, ShuffleNetV2, RegNet and YOLOv5 methods are employed to analyze the performance of DI-CNN algorithm. It is shown that the DI-CNN approach has a prediction accuracy that is 48.8% better than Transformer.

Surround-view cameras combined with image depth transformation to 3D feature space and fusion with point cloud features are highly regarded. The transformation of 2D features into 3D feature space by means of predefined sampling points and depth distribution happens throughout the scene, and this process generates a large number of redundant features. In addition, multimodal feature fusion unified in 3D space often happens in the previous step of the downstream task, ignoring the interactive fusion between different scales. To this end, we design a new framework, focusing on the design that can give 3D geometric perception information to images and unify them into voxel space to accomplish multi-scale interactive fusion, and we mitigate feature alignment between modal features by geometric relationships between voxel features. The method has two main designs. First, a Segmentation-guided Image View Transformation module is used to accurately transform the pixel region containing the object into a 3D pseudo-point voxel space with the help of a depth distribution. This allows subsequent feature fusion to be performed in a unified voxel feature. Secondly, a Voxel-centric Consistent Fusion module is used to alleviate the errors caused by depth estimation, as well as to achieve better feature fusion between unified modalities. Through extensive experiments on the KITTI and nuScenes datasets, we validate the effectiveness of our camera-LIDAR fusion method. Our proposed approach shows competitive performance on both datasets and outperforms state-of-the-art methods in certain classes of 3D object detection benchmarks. https://github.com/no-Name128/DLFusion [code release]

Multiobject tracking (MOT) technology can be deployed to the Internet of Things (IoT) devices to enhance the security and reliability of some video analysis applications, such as video surveillance and intelligent security system. However, since the IoT devices with limited computing capacity and storage, most existing MOT methods are difficult to deploy to IoT devices and exhibit poor tracking robustness in scenes with frequent occlusions, severe crowded, and scale variations. To alleviate the aforementioned issues, we propose a regression-based online MOT method. First, an object-aware embedding extraction module (OAEM) is designed to extract preliminary discriminative embedding of the object in the current frame. Then, an embedding aggregation module (EAM) is proposed to obtain high-quality aggregated embedding in temporal. Finally, the temporal embedding combined with the preliminary embedding extracted from the current frame to obtain a refined embedding feature for subsequent position prediction and association. Importantly, we achieve a beneficial interaction between embedding extraction, position prediction and association task. The proposed method does not suffer from significant memory consumption. Therefore, our method is a potential solution for intelligent video analysis on IoT devices. To evaluate the proposed method, we have conducted numerous experiments on the MOT16, MOT17, and MOT20 benchmark data sets. Results demonstrate that the proposed method can provide a more robust tracking performance compared to other optimal methods.