摘 要: 针对在边缘端算力较弱的设备上,动态手势检测模型较大、检测时间较长的问题,提出了以YOLOv5网络为基础的动态手势识别跟踪算法。首先在YOLOv5网络模型上添加自适应注意力模块输出大小相同的张量,添加特征增强模块弥补高层特征层的信息损失,提高特征金字塔的表示能力。其次使用Kalman滤波器预测机,形成有预测机制的动态手势识别跟踪算法。预测跟踪时,把短时的手部运动粗略地看成是一种恒定的运动,即在设定的非常短的时间内目标的状态是不变的,也就是相邻两帧图像状态差异不大,所以手势的运动模型采用的是等速度运动模型。利用大量正负样本训练模型,证明算法能实时地对13种动态手势进行识别和跟踪,识别准确率达到99.6%。 |
关键词: 改进YOLOv5;动态手势;Kalman;预测跟踪 |
中图分类号: TP391
文献标识码: A
|
基金项目: 2023年甘肃省高等学校创新基金项目(2023A-163);2022年甘肃省大学生创新创业项目(S202211807037X). |
|
Research on Dynamic Gesture Recognition and Tracking Algorithm Based on Improved YOLOv5 + Kalman |
ZHANG Ruimin, DU Shuqiang, ZHOU Xiuyuan
|
(College of Computer and Artificial Intelligence, Lanzhou Institute of Technology, Lanzhou 730050, China)
371247813@qq.com; 42787413@qq.com; 46652899@qq.com
|
Abstract: Aiming at the problems of large dynamic gesture detection models and long detection time on devices with weak computing power at the edge, this paper proposes a dynamic gesture recognition and tracking algorithm based on YOLOv5 network. Firstly, an adaptive attention module with the same output size tensor is added to the YOLOv5 network model and a feature enhancement module is added to compensate for the information loss in the high-level feature layer, so as to improve the representation ability of the feature pyramid. Then, the Kalman filter predictor is used to form a dynamic gesture recognition tracking algorithm with a prediction mechanism. When predicting the tracking, short-term hand movements are roughly regarded as a constant motion, meaning that the target's state remains unchanged for a set very short period of time. That is to say, there is not much difference in the states of adjacent frames of images. Therefore, the motion model of the gesture adopts an equal velocity motion model. The experiment utilizes a large number of positive and negative samples to train the model, proving that the proposed algorithm can recognize and track 13 dynamic gestures in real-time, with a recognition accuracy of 99.6% . |
Keywords: improved YOLOv5; dynamic gesture; Kalman; predictive tracking |