North China Institute of Science and Technology

为解决车内温度异常带来的安全隐患,提出一种基于人脸及背影检测的车内安全保障系统。采用Jetsonnano硬件结合OpenCV图像处理及yolov5目标检测算法实现。通过制作专门应用于车内场景的数据集来提高模型的鲁棒性。针对车内场景改进yolov5算法,采用Kmeans++算法生成性能更好的锚框、优化CBMA注意力模块提高算法对人脸及背影特征的关注度、EIOU损失函数解决预测框与真实框在长宽比相同且中心点重合时的定位问题。通过实验对比算法改进后检测的准确率(Precision)提高了4.6%,召回率(Recall)提高了1.3%,mAP_0.5提高了8.8%,mAP_0.5:0.95提高了4.4%。人脸及背影检测实验表明在车内人脸及背影遮挡情况下依然能够正确检测。通过车内实时温度监测,结合车内人脸及背影检测模型,实验表明通过Jetsonnano控制预警及车窗升降从而保障生命安全是可行的。

In order to solve the security risks caused by abnormal temperature inside the car, an in-vehicle safety and security system based on face and back shadow detection is proposed. It is realized by using Jetson nano hardware combined with OpenCV image processing and yolov5 target detection algorithm. The robustness of the model is improved by creating a dataset specifically applied to the in-vehicle scene. The yolov5 algorithm is improved for the car scene by using Kmeans++ algorithm to generate anchor frames with better performance, optimizing the CBMA attention module to improve the algorithm's focus on face and back features, and the EIOU loss function to solve the localization problem of the predicted frames and the real frames in the case that the aspect ratio is the same and the centroids are coincident. Through experimental comparison, the accuracy of the improved algorithm increases by 4.6%, the recall increases by 1.3%, the mAP_0.5 increases by 8.8%, and the mAP_0.5:0.95 increases by 4.4%. The face and back shadow detection experiments show that the face and back shadow can still be detected correctly in the case of face and back shadow occlusion in the car. Through the real-time temperature monitoring in the car, combined with the face and back shadow detection model, the experiment shows that it is feasible to control the warning and window lifting and lowering by Jetson nano to protect the life safety.

中央高校基本科研业务费资助项目(3142015024);河北省物联网监控工程技术研究中心项目基金(3142016020)

张超,顾涛.基于人脸及背影检测的车内安全保障系统研究[J].华北科技学院学报,2024,21(3):50-58

ZHANG Chao, GU Tao. Research on in-vehicle safety and security system based on face and back detection[J].Journal of North China Institute of Science and Technology,2024,21(3):50-58