基于复杂环境的人员跌倒检测算法OA
Pedestrian Fall Detection Algorithm Based on Complex Environments
针对复杂环境中存在目标遮挡以及小目标等问题,文中提出了一种基于YOLOv7(You Only Look Once ver-sion 7)模型改进的人员跌倒检测算法.改进模型引入SPDConv(Space-to-Depth Convolution)与InceptionNeXt相结合的新模块SPINeXt,以此充分提取特征信息,从而提升在遮挡环境下的跌倒检测性能.将归一化的Wasserstein距离损耗按比例添加到原始CIoU(Complete Intersection over Union),增强了原YOLOv7模型对小目标位置偏差的敏感度.采用Mosaic和Mixup数据增强技术进一步提升原YOLOv7模型的鲁棒性,从而获得更准确的检测结果.实验结果表明,改进模型在复杂环境下的跌倒检测精确率为0.956,优于其他网络模型,验证了所提模型在复杂环境下进行人员跌倒检测的综合性能最优.
In view fo the problems such as target occlusion and small targets in complex environments,an im-proved personnel fall detection algorithm based on the YOLOv7(You Only Look Once version 7)model is proposed.The improved model introduces a new module,SPINeXt,which combines SPDConv(Space-to-Depth Convolution)with InceptionNeXt,to fully extract feature information and thereby enhance the fall detection performance in oc-clented environments.The normalized Wasserstein distance loss was proportionally added to the original CIoU(Com-plete Intersection over Union),enhancing the sensitivity of the original YOLOv7 model to the position deviation of small targets.The Mosaic and Mixup data augmentation techniques are adopted to further enhance the robustness of the original YOLOv7 model,thereby obtaining more accurate detection results.The experimental results show that the accuracy rate of fall detection of the improved model in complex environments is 0.956,which is superior to oth-er network models,verifying that the proposed model has the best comprehensive performance in personnel fall de-tection in complex environments.
王菲;张运楚;张汉元
山东建筑大学 信息与电气工程学院,山东 济南 250101山东建筑大学 信息与电气工程学院,山东 济南 250101||山东省智慧建筑与建筑节能重点实验室,山东 济南 250101山东建筑大学 信息与电气工程学院,山东 济南 250101||山东省智慧建筑与建筑节能重点实验室,山东 济南 250101
信息技术与安全科学
跌倒检测深度学习图像处理神经网络YOLOv7损失函数数据增强目标检测
fall detectiondeep learningimage processingneural networkYOLOv7loss functiondata enhancementtarget detection
《电子科技》 2026 (10)
34-41,8
国家自然科学基金(62003191) National Natural Science Foundation of China(62003191)
评论