Intelligent Transportation Systems (ITS) rely on high-precision road networks, which are particularly scarce in indoor parking. Existing methods depend on expensive hardware (e.g., LiDAR) or manual mapping, both of which are costly and inefficient. The rise of the Internet of Things (IoT) has enabled large-scale data collection and connectivity, offering new opportunities for automated road network extraction. OpenStreetMap (OSM), as a crowdsourced IoT-driven platform, provides multi-layer geospatial data, including the Road Network Layer (RNL), Lane Boundary Layer (LBL), and Turn Sign Layer (TSL). However, OSM data often suffers from incompleteness and noisy connectivity, affecting the continuity and accuracy of road networks. This paper introduces OSM2Net, a novel framework designed to extract road networks from individual layers and leverage multi-layer data to construct directed road networks. Specifically, OSM2Net rasterizes noisy OSM data into bitmaps for image processing and multi-layer fusion. By leveraging the topology relationship between lane boundaries and road networks, a Lane-Road Map Generator (LRMG) creates a simulated dataset for training. Then, utilizing the simulated dataset, a Lane2Net model is designed to extract road networks from sparse lane boundary images. The framework then vectorizes bitmaps into a lightweight, undirected road network and refines it into a directed network by extracting and matching turn sign information. Experimental results show that Lane2Net achieves Intersection over Union (IoU) of 93% and 92% using simulated and real-world datasets, respectively. Extensive experiments on real-world datasets confirm that OSM2Net delivers robust completeness and high-quality road network extraction.

Non-line-of-sight (NLOS) identification is the key technique to improve the accuracy of the channel impulse response (CIR) based ultrawideband (UWB) positioning system. However, most existing NLOS identification approaches are tailored to static environments and often encounter difficulties in dynamic settings with both temporal and spatial variations, particularly when dealing with limited and hard samples. This paper introduces a hard sample meta-learning (HSML) approach to address the issues of NLOS identification across different scenarios and domains. HSML includes two phases: a hard sample meta-training phase and a fine-grained meta-testing phase. During the meta-training phase, we train a two-loop learning network using CIR from multiple scenarios (tasks). The inner loop focuses on learning task-specific features, while the outer loop captures cross-task generalization properties using a cross-entropy loss. Hard samples are identified based on estimated residuals for each task, and a new dataset is created, consisting of both hard samples and samples with small residuals. To improve the robustness against hard samples, we implement a residual-corrected focal loss, which is used to retrain the network on this new dataset. In the fine-grained meta-testing phase, we apply a filtering mechanism based on the tendency of estimated residuals during fine-tuning. This mitigates the risk of poor performance caused by anomalous samples. We validate the effectiveness and robustness of the proposed HSML method using two datasets containing multiple real-world scenarios. Our experimental results demonstrate that HSML outperforms existing models in terms of identification accuracy, robustness and generalization performance.

Efficient trajectory planning in Autonomous Valet Parking (AVP) remains challenging due to multiple vehicle interactions and environmental complexities. Existing single-agent Reinforcement Learning (RL) approaches face challenges in balancing complexity, convergence, and knowledge efficiency, often resulting in increased collisions and longer travel times. To address these issues, this paper proposes a Bio-inspired Evolutionary Reinforcement Learning (BERL) framework for multi-vehicle collaborative trajectory planning, where each vehicle is modeled as a Fusion Architecture for Learning and Cognition Network (FALCON) agent based on Adaptive Resonance Theory (ART). The BERL framework comprises three core modules: 1)Meme Reinforcement Learning (MRL), which enables agents to learn independently and adapt to changing environments; 2)Expert-Guided Evolutionary Learning (EGEL), which facilitates knowledge transfer from expert agents to less experienced ones, enhancing coordination; and 3)Integrated Forgetting and Memory Optimization (IFMO), which optimizes memory use and reduces algorithm complexity. Additionally, the BERL framework supports model and sensor quality heterogeneity in the multi-vehicle trajectory planning scenario. Finally, we build an AVP Simulation (AVPS) platform to validate the performance of the proposed framework. Comprehensive simulation results demonstrate that the BERL framework improves success rate and parking efficiency by at least 15.7% and 16.7%, respectively, as compared to state-of-the-art algorithms. Additionally, the proposed IFMO module reduces the number of memes in the FALCON agent by 30.2% while maintaining stable performance.

Crowdsourcing is an effective method to alleviate the burden of conducting a site-survey procedure for localization tasks. However, crowdsourced data is typically inaccurately and scarcely annotated, rendering accurate localization a rather challenging problem. To alleviate this problem, we propose VRLoc, a deep reinforcement learning (DRL)-based unsupervised wireless localization framework using crowdsourced trajectory data. The proposed VRLoc primarily encompasses three components, i.e., a robust K-means (RKM) clustering method for generating a series of virtual reference points (VRPs), DRL for determining the physical layout for VRPs, and online localization based on VRPs. Specifically, the proposed RKM method employs a density-based approach for the initialization of cluster centers, rather than the commonly used random solution, yielding repeatable and reliable VRP generation results. To accurately determine the physical locations for VRPs, we develop a modified soft actor-critic (SAC)- based VRP layout method with multiple objectives, i.e., the connection topology among VRPs, the floor-plan information, and the near-field condition. Then, we effectively predict locations of target users by utilizing classification models to match the online collected samples with the VRPs annotated by physical locations. The proposed framework is advantageous in achieving high-accuracy unsupervised localization, with the VRPs bridging the unlabeled crowdsourced data and physical location space. Both experimental and simulation results demonstrate the effectiveness and superiority of the proposed VRLoc framework as an accurate and practical solution for unsupervised localization.

The heavy burden of fingerprint collection and annotation has become one of the biggest bottlenecks in wireless indoor positioning, particularly in the context of the Internet of Things (IoT). Fortunately, crowdsourcing can be leveraged to alleviate the fingerprint collection burden by harnessing the collective intelligence of crowdsourcing users. However, it is rather difficult to acquire an accurate positioning model based solely on training unlabeled crowdsourcing data. To overcome this problem, we propose a novel positioning model called ENvironment Aware Positioning (ENAP), utilizing unlabeled crowdsourcing trace data. The proposed ENAP mainly consists of three steps, i.e., transforming the unlabeled crowdsourcing trace data into a cluster space, mapping the cluster space into the positioning space, and continuously updates the positioning model in an unsupervised manner. To enhance the performance and robustness against device heterogeneity of crowdsourcing users, we propose a novel clustering scheme for space transformation by adaptively fusing multiple signal features. Then, to ensure long-term positioning stability and continual environmental aware capability, we incorporate a dynamic replay memory into ENAP that enables the unsupervised online updating of positioning models, distinguishing our proposal from most existing positioning models. Simulation and experimental results demonstrate the effectiveness and superiority of the proposed ENAP approach as a practical and efficient solution for wireless indoor positioning in the IoT era.