摘要
为了探索一种新的校园安全监测模式,研究提出基于图关联注意力特征的跨模态行人检测方法。该方法将多模态数据进行融合,并利用注意力机制和掩模引导来增强模型的泛化能力。室内搜索模式下,当排名Rank为20时,相比于仅输入完整图像,前景图像和完整图像同时作为输入,可将网络的识别率提升6.7%。相比于单独的通道注意力模块和空间注意力模块,两者同时使用可将网络的识别率提升3.1%。改进后的校园安全平台的行人检测的识别率高于93%,行人轨迹预测的准确率为97.8%。相比于改进前的校园安全平台,改进后的校园安全平台的整体性能满意度评价提升了31.2%。研究提出的跨模态行人检测方法,提高了校园内行人检测的准确性和可靠性。这有助于保障校园安全,预防和减少安全事故的发生。
In order to explore a new campus safety monitoring mode,a cross modal pedestrian detection method based on graph associated attention features is proposed.This method integrates multimodal data and utilizes attention mechanisms and mask guidance to enhance the model's generalization ability.In indoor search mode,when the Rank is 20,compared to only inputting complete images,using both foreground and complete images as inputs can improve the recognition rate of the network by 6.7%.Compared to separate channel attention modules and spatial attention modules,the simultaneous use of both can improve the recognition rate of the network by 3.1%.The recognition rate of pedestrian detection on the improved campus safety platform is higher than 93%,and the accuracy of pedestrian trajectory prediction is 97.8%.Compared to the pre improved campus security platform,the overall performance satisfaction evaluation of the improved campus security platform has increased by 31.2%.The cross modal pedestrian detection method proposed in the study has improved the accuracy and reliability of pedestrian detection on campus.This helps to ensure campus safety,prevent and reduce the occurrence of safety accidents.
作者
黄洪松
梁红娥
HUANG Hongsong;LIANG Hong'e(Xi'an Siyuan College,Xi’an 710038,China)
出处
《自动化与仪器仪表》
2024年第11期154-157,162,共5页
Automation & Instrumentation
基金
陕西省“十四五”教育科学规划2023年度课题《三全育人视域下高校朋辈心理互助体系的构建与路径研究》阶段成果之一(SGH23Y2886)。
关键词
掩模图像
注意力机制
特征提取
跨模态
行人检测
校园安全
mask image
attention mechanism
feature extraction
cross modal
pedestrian detection
campus security
