计算机科学
拥挤感测
感知
人工智能
移动电话技术
流量(数学)
实时计算
数据建模
移动计算
计算机网络
流量(计算机网络)
移动设备
数据挖掘
弹道
无线传感器网络
流量网络
计算机视觉
作者
Yinxin Bao,Qinqin Shen,Yang Cao,Quan Shi
标识
DOI:10.1109/jiot.2026.3685425
摘要
Mobile Crowdsensing (MCS) has recently emerged as an effective paradigm for large-scale traffic state estimation and prediction in Intelligent Transportation System (ITS). By integrating data from vehicular sensors and mobile devices, MCS enables real-time and fine-grained traffic monitoring. However, MCS data are often sparse, incomplete, and noisy due to limited participant coverage and unstable sensing environments, which degrades the accuracy and robustness of traffic flow prediction. To address these challenges, this paper proposes a Neighborhood Perception Enhanced Multi-Factor Causal Gated Network (NP-MCGN) for traffic flow prediction based on sparse MCS data. The model incorporates a Neighborhood Perception (NP) module that integrates graph convolution with neighbor-weighted estimation to repair missing data, and a Multi-Factor Spatial Encoding (MFSE) module that fuses distance, correlation, and dynamic structure graphs via graph attention and Mamba state space modeling to capture multi-source spatial dependencies. Furthermore, a Causal Temporal Encoding (CTE) module couples causal convolution with the Mamba architecture to model both short-term fluctuations and long-term temporal evolution, while a Gated Recurrent Decoding (GRD) module dynamically fuses spatial–temporal representations for multi-step forecasting. Extensive experiments on seven real-world MCS-based datasets, including PEMS03, PEMS04, PEMS07, PEMS08, SZ-TAXI, METR-LA, and PEMS-BAY, demonstrate that NP-MCGN consistently outperforms state-of-the-art models under both complete and missing data conditions, verifying its robustness and applicability in realistic MCS environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI