基于机器学习的人体姿态估计模型研究

A Study on the Human Attitude Estimation Model Based on Machine Learning

设计了一种新型姿态估计算法,支持同步执行多个任务,在行人姿态估计中可以达到较好的效果。网络中划分为主网络、子网络两大部分,前者采用的主要是Darknet53网络,用于实现特征提取的功能;后者划分为关键点检测以及行人检测两大部分,分别基于特征金字塔、YOLOv3,将网络融合之后具备了显著的优势,继承了各个网络的优点,在保证姿态识别精度的同时提高了效率,减少了预测的时间,适合于应用到实时性要求较高的场景中。通过姿态残差网络来学习先前获得的姿态信息,解决了由于行人重叠或者遮挡等导致的图像模糊问题,依然保持了较高的识别精度,相对于其他的方法达到更高的准确性。基于与其他算法(CMU-Pose、Newel)进行对比的结果可知,设计的算法在COCO数据集中的预测精度达到了67.2%,效率为20 FPS,相对于其他算法无论在效率还是在精度上均显著提高,这表明提出的算法在保证准确率的同时具有实时性处理效果。

A multi-task-based pedestrian pose estimation algorithm is designed to combines Darknet53 network with YOLOv3-based target detection sub-network and feature pyramid-based key detection sub-network to realize multi-task detection and shorten the running time of the algorithm,the problem in the image to improve the accuracy of pedestrian pose estimation.The experimental results show that the algorithm proposed in this paper and the similar algorithms CMU-Pose and Newel algorithms compared,on the COCO test set,the running speed reached 20 FPS,improved by nearly 2 times,the average accuracy reached 67.2%,improved by about 2%,indicating that the proposed algorithm has real-time processing effect while ensuring the accuracy.