Shangwei Xie

Xidian University
1 previous affiliation
2Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Shangwei Xie focuses on computer vision and video anomaly detection (VAD), with a specific emphasis on the identification and analysis of crowd-level abnormal behaviors (CABs). This research addresses the technical challenges of distinguishing collective motion anomalies—such as counter-flows, crowd turbulence, and irregular density shifts—from individual-level anomalies like fighting or running. The work is primarily concerned with the safety of high-density public spaces, where local observations may appear normal while the global or meso-scale interactions indicate significant risk.

Research Trajectory

The available evidence indicates a research trajectory centered on the development and refinement of motion consistency learning frameworks. Between 2022 and 2024, this work evolved from the initial conceptualization of multi-scale detection methods to the more complex task of quantifying risk levels within crowd events.

In the early stages of this period (2022–2023), the research focused on the fundamental problem of scale in crowd dynamics. Traditional VAD techniques often fail to detect CABs because these behaviors do not exhibit salient differences from normal patterns when observed at a local level. Xie participated in the development of a multi-scale motion consistency learning approach designed to capture the complex interactions of individuals across different spatial and temporal resolutions. This work established that detecting crowd disasters requires an understanding of how individual movements aggregate into larger, potentially hazardous patterns.

By 2024, the focus shifted toward the quantification of these behaviors. While detection identifies the presence of an anomaly, quantification is necessary for practical surveillance applications to assess the actual risk level of a given event. This transition involved improving the existing motion learning frameworks to not only flag abnormal motion but also provide a measurable assessment of its severity. This progression suggests a move from purely diagnostic computer vision toward more prescriptive tools for public safety and crowd management.

Representative Work

The following papers represent the core of this research trajectory, moving from the detection of crowd anomalies to their quantification.

In Crowd-level Abnormal Behavior Detection via Multi-scale Motion Consistency Learning (and its subsequent publication in AAAI 2023), Xie participated as a co-author in the development of a framework designed to address the limitations of individual-level anomaly detection. The research posits that crowd-level anomalies are often the precursors to disasters and are characterized by subtle, large-scale motion inconsistencies. The proposed method utilizes multi-scale learning to identify these patterns, which are otherwise invisible to algorithms focused on local pixel-level or individual-object changes.

In Detecting and Quantifying Crowd-Level Abnormal Behaviors in Crowd Events, Xie served as the second author, contributing to an expanded version of the motion learning framework. This work addresses the challenge of risk assessment by introducing methods to quantify the level of abnormality in crowd motion. This is particularly relevant for video surveillance, where security personnel require more than a binary alert; they need data on the scale and intensity of the behavior to prioritize interventions. The paper refines the multi-scale motion consistency approach to provide these quantitative measures, representing a significant advancement in the utility of VAD for real-world safety monitoring.

The current evidence for this researcher is focused entirely on the domain of crowd dynamics and video analysis. While the work demonstrates a clear progression in technical complexity and application, the scope of the available corpus is limited to this specific sub-field of computer vision.

Profile updated

See something wrong? Report a mistake