摘要
分析大学生在校行为模式与活动规律是提升高校精细化管理、保障校园安全、促进学生心理健康的重要基础。随着以视频监控为核心的综合安防系统在高校的广泛部署,海量带有时空标记的学生轨迹数据为行为建模提供了新机遇。然而,如何从非结构化的视频流中提取高质量轨迹,并实现跨场景、长周期的行为预测与异常感知,仍是亟待解决的问题。针对该问题,文章提出一种面向校园场景的学生日常行为预测框架(campus behavior prediction framework,CBPF),该框架由结构化轨迹生成、时空语义嵌入、时空预测基础模型(spatio-temporal prediction foundation model,STPFM)和预测-异常联合优化4个模块组成,其核心STPFM基于时空注意力机制(spatio-temporal attention,STA),能够自适应建模学生在多区域间的迁移规律与时间周期性。以某高校681名学生在30 d内的79843条轨迹为例,对比分析不同模型在行为趋势预测和异常感知的评价指标,结果表明,在同一框架下STPFM在行为趋势预测任务中的平均绝对误差(mean absolute error,MAE)比长短期记忆网络(long short-term memory,LSTM)等基线模型降低了14.2%,异常检测的F 1值提升了10.3%。研究结果验证了所提模型在捕捉长期行为规律与识别潜在风险方面的有效性。
Analyzing on-campus behavioral patterns and activity routines of university students is essential for enhancing meticulous management in universities,ensuring campus security,and promoting student psychological well-being.The widespread deployment of integrated security systems based on video surveillance in universities has generated large-scale student trajectory data with spatio-temporal tags,offering new opportunities for behavioral modeling.However,extracting high-quality trajectories from unstructured video streams and achieving cross-scene,long-term behavior prediction and anomaly perception remain challenging.To address these issues,this paper proposes a campus behavior prediction framework(CBPF),which comprises four modules:structured trajectory generation,spatio-temporal semantic embedding,spatio-temporal prediction foundation model(STPFM),and joint prediction-anomaly optimization.The STPFM leverages a spatio-temporal attention(STA)mechanism to adaptively model students’transition patterns across multiple regions and their temporal periodicities.Experiments were conducted on a dataset of over 79843 trajectories from 681 students at a university over 30 days to comparatively analyze the evaluation metrics of different models in behavioral trend prediction and anomaly perception.Results demonstrate that within the unified framework,STPFM reduces mean absolute error(MAE)by 14.2%in behavioral trend prediction compared to baseline models such as long short-term memory(LSTM),and improves the anomaly detection F 1-score by 10.3%.These findings validate the effectiveness of the proposed model in capturing long-term behavioral regularities and identifying potential risks.
作者
于长伟
周立
闫艾婧
YU Changwei;ZHOU Li;YAN Aijing(Party and Administration Office,Hefei University of Technology,Hefei 230009,China;Party Committee Security Department,Hefei University of Technology,Hefei 230009,China;Hangzhou Aiforia Robot Co.,Ltd.,Hangzhou 311200,China)
出处
《合肥工业大学学报(自然科学版)》
北大核心
2025年第12期1671-1677,共7页
Journal of Hefei University of Technology:Natural Science
基金
国家自然科学基金资助项目(62476077)。
