Vehicle and pedestrian detection is one of the critical tasks in autonomous driving. Since heterogeneous techniques have been proposed, the selection of a detection system with an appropriate balance among detection accuracy, speed and memory consumption for a specific task has become very challenging. To deal with this issue and to provide guidance for model selection, this paper analyzes several mainstream object detection architectures, including Faster R-CNN, R-FCN, and SSD, along with several typical feature extractors, such as ResNet50, ResNet101, MobileNet_V1, MobileNet_V2, Inception_V2 and Inception_ResNet_V2. By conducting extensive experiments using the KITTI benchmark, which is a commonly used street dataset, we demonstrate that Faster R-CNN ResNet50 obtains the best average precision (AP) (58%) for vehicle and pedestrian detection, with a speed of 8.6 FPS. Faster R-CNN Inception_V2 performs best for detecting cars and detecting pedestrians respectively (74.5% and 47.3%). ResNet101 consumes the highest memory (9907 MB) and has the largest number of parameters (64.42 millions), and Inception_ResNet_V2 is the slowest model (3.05 FPS). SSD MobileNet_V2 is the fastest model (70 FPS), and SSD MobileNet_V1 is the lightest model in terms of memory usage (875 MB), both of which are suitable for applications on mobile and embedded devices.

Acquiring accurate and reliable pedestrian trajectories is essential for providing indoor location-based services. Although a foot-mounted inertial navigation system (Foot-INS) can acquire pedestrian trajectories in multistory buildings, it will inevitably encounter heading divergence because the constraint information is not always valid. Therefore, we proposed an accurate and convenient postprocessing indoor pedestrian positioning system (IPPS) to acquire pedestrian trajectories in multistory buildings such as shopping malls. Based on the hypotheses that the start and end points of the pedestrian trajectories on a single floor were closed, and the horizontal position of the closing point on each floor was identical. Therefore, in the single floor of multistory buildings, we use the closing point to control the trajectory drift error caused by the Foot-INS, and use a smoothing algorithm to reasonably distribute the drift error to the entire trajectory. Heading divergence is unavoidable with the Foot-INS, result in the pedestrian trajectories acquired on different floors were rotationally offset. Because pedestrian trajectories can epitomize the building orientation and provide an opportunity to align those trajectories on multistory buildings, an algorithm was proposed to match the trajectories acquired on different floors. A hybrid simulation experiment was conducted using an accurate reference object to evaluate the positioning performance of the proposed IPPS. The effectiveness of acquiring pedestrian trajectories was also confirmed by various experimental tests in a large shopping mall. The study findings suggest that the proposed IPPS is self-contained, low cost, and has the potential for large-scale applications.

This paper proposes an Equivariant Filtering (EqF) framework for the inertial-integrated state estimation. As the kinematic system of the inertial-integrated navigation can be naturally modeled on the matrix Lie group SE₂(3), the symmetry of the Lie group can be exploited to design an equivariant filter which extends the invariant extended Kalman filtering on the group-affine system and overcomes the inconsitency issue of the traditional extend Kalman filter. We firstly formulate the inertial-integrated dynamics as the group-affine systems. Then, we prove the left equivariant properties of the inertial-integrated dynamics. Finally, we design an equivariant filtering framework on the earth-centered earth-fixed frame and the local geodetic navigation frame. The experiments show the superiority of the proposed filters when confronting large misalignment angles in Global Navigation Satellite Navigation (GNSS)/Inertial Navigation System (INS) loosely integrated navigation experiments.

Spectrum sharing in cellular vehicle-to-everything (C-V2X) has been conceived as a promising solution to improve spectrum efficiency. However, the co-channel interference incurred with it may cause severe performance degradation to vehicular links. Thereby, radio resource management (RRM) is motivated and designed to ensure communication reliability and increase system capacity. One challenge is that RRM involves channel allocation and power control, which are tightly coupled and hard to optimize simultaneously. Another challenge for this is the difficulty adapting centralized RRM schemes, requiring global channel state information (CSI) and causing high signaling overhead. To tackle these challenges, we propose the hybrid centralized-distributed RRM scheme and the distributed RRM scheme. Specifically, we prove a decoupling method that provides a theoretical lower bound so that channel allocation and power control can be optimized independently. Given the decoupling method, the hybrid centralized-distributed RRM scheme is based on graph matching and reinforcement learning (GMRL) to maximize system capacity and guarantee reliability requirements. Further, to decrease computation complexity and signaling overhead, the distributed RRM scheme that only requires local CSI with hybrid-framework reinforcement learning (HFRL) is exploited. Finally, both schemes are numerically evaluated through experiments and outperform other deep Q-network (DQN)-based schemes.

With the rapid development of intelligent driving vehicles, multi-task visual perception based on deep learning emerges as a key technological pathway toward safe vehicle navigation in real traffic scenarios. However, due to the high-precision and high-efficiency requirements of intelligent driving vehicles in practical driving environments, multi-task visual perception remains a challenging task. Existing methods typically adopt effective multi-task learning networks to concurrently handle multiple tasks. Despite the fact that they obtain remarkable achievements, better performance can be achieved through tackling existing problems like underutilized high-resolution features and underexploited non-local contextual dependencies. In this work, we propose YOLOPv3, an efficient anchor-based multi-task visual perception network capable of handling traffic object detection, drivable area segmentation, and lane detection simultaneously. Compared to prior works, we make essential improvements. On the one hand, we propose architecture enhancements that can utilize multi-scale high-resolution features and non-local contextual dependencies for improving network performance. On the other hand, we propose optimization improvements aiming at enhancing network training, enabling our YOLOPv3 to achieve optimal performance via straightforward end-to-end training. The experimental results on the BDD100K dataset demonstrate that YOLOPv3 sets a new state of the art (SOTA): 96.9% recall and 84.3% mAP50 in traffic object detection, 93.2% mIoU in drivable area segmentation, and 88.3% accuracy and 28.0% IoU in lane detection. In addition, YOLOPv3 maintains competitive inference speed against the lightweight YOLOP. Thus, YOLOPv3 stands as a robust solution for handling multi-task visual perception problems. The code and trained models have been released on GitHub.

There is an increasing emphasis on achieving high accuracy and robustness in SLAM systems. The traditional visual-inertial simultaneous localization and mapping (VI-SLAM) system often struggles with stability under low light or dynamic illumination, leading to the potential loss of trajectory tracking. High accuracy and robustness are crucial for ensuring the long-term stability and reliable localization performance of SLAM systems. To address the challenges of improving robustness and accuracy in VI-SLAM, this article proposes SuperVINS, a real-time VI-SLAM framework designed for challenging imaging conditions. In contrast to geometric modeling, deep learning-based features are capable of fully leveraging the implicit information present in images, which is often not captured by geometric features. Therefore, SuperVINS, developed as an enhancement of VINS-Fusion, integrates the deep learning neural network model SuperPoint for feature point extraction and loop closure detection. At the same time, a deep learning neural network LightGlue model for associating feature points is integrated into front-end feature matching. We employed the RANSAC algorithm for matching enhancement to improve robustness against outliers. In addition, SuperVINS enables flexibly integrated environment-specific SuperPoint bag-of-words (BOWs) models for improved loop closure detection. The system enables real-time localization and mapping. Experimental validation on the well-known EuRoC and UMA-VI datasets demonstrates that SuperVINS achieves comparable accuracy and robustness to other state-of-the-art VI-SLAM systems, particularly in the most challenging sequences. This article analyzes the advantages of SuperVINS in terms of accuracy, real-time performance, and robustness. To foster knowledge exchange in the field, we have publicly released the code associated with this article. The code is available at https://github.com/luohongk/SuperVINS.