基于改进YOLOv7的室内摔倒行为检测
2024,32(12):35-42
摘要：针对室内监控视频中老年人摔倒行为的检测问题，提出一种基于改进 YOLOv7 网络模型的实时摔倒行为检测算法。基于YOLOv7的目标检测模型传统使用跨步卷积来实现下采样特征，但这可能会使目标信息的特征模糊。为了解决这个问题，引入了新的下采样模块——鲁棒特征下采样，以改善下采样过程中目标信息特征的清晰度。此外，通过在网络的 concat 部分引入 CoordAttention 注意力机制，可更好地融合拼接后的特征图。实验结果表明，改进后的YOLOv7模型在摔倒行为检测方面具有较高的准确率和检测性能，准确率达到98.88%，mAP50值达到98.83%，mAP50-95值达到74.12%。这意味着该算法可以准确地检测老年人的摔倒行为，家人能够及时地发现，以便及时采取必要的救助措施。
关键词：YOLOv7网络模型;摔倒检测;下采样;鲁棒特征下采样;CoordAttention注意力机制
Indoor fall behavior detection algorithm based on improved YOLOv7
Abstract：Aiming at the problem of fall behavior detection in indoor surveillance video, a real-time fall behavior detection algorithm based on improved YOLOv7 network model was proposed. The target detection model based on YOLOv7 traditionally uses strided convolution to realize the feature of downsampling, but this may make the feature of the target information fuzzy. To solve this problem, a new downsampling module, robust feature downsampling, is introduced to improve the clarity of target information features during downsampling. In addition, by introducing CoordAttention attention mechanism in the concat portion of the network, the spliced feature graphs can be better merged. The experimental results show that the improved YOLOv7 model has a high accuracy and detection performance in fall behavior detection, with the accuracy reaching 98.88%, the mAP50 value reaching 98.83%, and the mAP50-95 value reaching 74.12%.This means that the algorithm can accurately detect the fall behavior of the elderly, and the family can find it in time so that the necessary rescue measures can be taken in time.
Key words：YOLOv7 network model;Fall detection; Downsampling;Robust feature downsampling; CoordAttention attention mechanism
收稿日期：2023-11-01
基金项目：空
