Chenxi Lin

Dr. Chenxi Lin

orcid

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 July 2024 in Expert Systems with Applications
Verified
Journal Article
Closed access
CLChenxi Lin
Chenxi Lin
XHXiaojian Hu
Xiaojian Hu
YZYunfei Zhan
Yunfei Zhan
XHXiatong Hao
Xiatong Hao

17 June 2025 in IEEE Transactions on Intelligent Transportation Systems
Verified
Journal Article
Closed access
CLChenxi Lin
Chenxi Lin
XHXiaojian Hu
Xiaojian Hu
Object counting in congested traffic scenes is an important component of traffic perception, facilitating urban traffic management and public transportation capacity optimization. Vision Transformer (ViT) models have achieved remarkable performance in object counting tasks due to their ability to capture long-range dependencies among diverse image blocks. In this paper, we observe that most existing ViT-based models typically employ fixed-size attention, which is insufficient to address the significant scale variation present in traffic scenes. To overcome this issue, we propose an effective Dynamic Multi-scale Local Vision Transformer (DMLViT) which introduces a priori multi-scale representation to effectively handle scale variation. Specifically, we carry out self-attention on multiple windows around each query token and dynamically fused these attention features to model multi-range spatial dependencies. Two dynamic aggregation strategies are designed to assign corresponding weights to different window branches along the channel or spatial dimensions. Furthermore, as the original feed-forward network (FFN) has limitation in model local information, we develop an efficient local FFN to better capture local dense information. Extensive experiments on four crowd counting datasets and three vehicle counting datasets demonstrate that the proposed DMLViT exhibits promising performance compared to state-of-the-art methods.

27 January 2025 in Sensors
Verified
Journal Article
Open access
PDF file
CJCheng Jian
Cheng Jian
CLChenxi Lin
Chenxi Lin
XHXiaojian Hu
Xiaojian Hu
JLJian Lu
Jian Lu
Traffic congestion detection in surveillance video is crucial for road traffic condition monitoring and improving traffic operation efficiency. Currently, traffic congestion is often characterized through traffic density, which is obtained by detecting vehicles or using holistic mapping methods. However, these traditional methods are not effective in dealing with the vehicle scale variation in surveillance video. This prompts us to explore density-map-based traffic density detection methods. Considering the dynamic characteristics of traffic flow, relying solely on the spatial feature of traffic density is overly limiting. To address these limitations, we propose a multi-task framework that simultaneously estimates traffic density and dynamic traffic congestion. Specifically, we firstly propose a Selective Scale-Aware Network (SSANet) to generate a traffic density map. Secondly, we directly generate a static congestion level from a traffic density map through a linear layer, which can characterize the spatial occupancy of traffic congestion in each frame. In order to further describe dynamic congestion, we simultaneously consider the dynamic characteristics of traffic flow, using the overall traffic flow velocity integrated with static congestion estimation for a dynamic assessment of congestion. On the collected dataset, our method achieves state-of-the-art results on both congestion detection and density estimation task. SSANet also obtains 99.21% accuracy on the UCSD traffic flow classification dataset, which outperforms other state-of-the-art methods.

27 January 2026 in IEEE Transactions on Intelligent Transportation Systems
Verified
Journal Article
Closed access
CLChenxi Lin
Chenxi Lin
XHXiaojian Hu
Xiaojian Hu
Ensuring safety in autonomous driving requires accurate pedestrian intention prediction in complex urban environments. Nevertheless, pedestrian motion is inherently influenced by environmental factors and visual appearance, which poses a significant challenge. Although existing multi-modal methods integrate motion data with both global and local visual cues during training, applying the same model structure for inference results in high computational cost and reduces real-time performance. To address this limitation, we propose the Global-Local Context Co-learning (GLCCL) framework, which leverages both local and global context during training for collaborative learning while relying solely on lightweight motion data for efficient real-time inference. Our framework integrates three key components: a Cross-modal Query Module (CQM) for modeling vehicle-pedestrian motion feature, a Local Perception Module (LPM) for extracting local pedestrian features, and a Global Guidance Module (GGM) for encoding global traffic environment descriptions. Collaboratively training these modules enables our framework to enrich motion representations with global-local contextual information. We further visualize the CQM’s cross-attention matrices to interpret the interactions of multi-modal motion features. Experiments on the public PIE and JAAD benchmarks demonstrate that, compared with state-of-the-art multi-modal models, GLCCL achieves competitive performance while offering superior real-time capabilities and lower computational cost. Extensive ablation studies confirm the framework’s compatibility with diverse sequence-learning methods.

15 July 2025 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
XHXiaojian Hu
Xiaojian Hu
CLChenxi Lin
Chenxi Lin
Accurate measurement of crowd density plays a critical role in public safety and intelligent transportation systems. In practical scenarios, cameras and embedded devices serve as essential instruments for real-time measurement. However, the discrepancy between the heavy network architecture frequently utilized for high-precision counting tasks and the constrained computational resource inherent in edge device has emerged as a significant challenge. In order to address this challenge, we propose an effective encoder-decoder based network, named lightweight multiscale perception and global context guidance network. Specifically, lightweight MobileOne network is applied as backbone for efficient feature extraction. Adaptive large kernel attention module is proposed for multiscale spatial hierarchy relationship extraction. Global context guidance module is designed for the fusion of multi-level features, allowing the extraction of effective global information from high-level feature to propagate to low-level feature. Experiments on four crowd counting datasets demonstrate that the proposed model is able to achieve comparable counting performance while maintaining small model size and low inference latency. The proposed model is deployed on Nvidia Jetson Nano and applied in edge density estimation of non-motorized transportation participants in Nanjing, providing important data foundation for traffic signal timing optimization.

1 August 2025 in Journal of Transportation Engineering, Part A: Systems
Verified
Journal Article
Closed access
XYXunming Yuan
Xunming Yuan
QCQian Chen
Qian Chen
JLJia Li
Jia Li
SGShuyan Gong
Shuyan Gong
CLChenxi Lin
Chenxi Lin
XHXiaojian Hu
Xiaojian Hu