我可能错了，但我认为您可以通过将视频中的整个挥动动作标记为正面来训练单个离散分类器。在训练中，分类器从诸如关节位置之类的弱特征中学习，但也从其衍生物（联合
速度和加速度）以及神秘计算的关节力，扭矩和力量中学习。因此，分类器可以得出结论：动态特征足以积极地检测动态运动，例如挥动并忽略静态
位置特征。我会确保在你的捕获数据中有一些手的实例位于与挥动相同的位置，但实际上并没有挥动。然后不要将这些实例标记为挥动的正面例子。

I might be wrong, but I would think that you can train a single discrete classifier by tagging the entire waving motion in your video as positive. In training the classifier learns from weak features such as joint position, but also its derivatives (joint velocity and acceleration) as well as mysteriously calculated joint forces, torques and powers. So the classifier may come to the conclusion that the dynamic features are sufficient to positively detect a dynamic motion such as waving and ignore the static position features. I would make sure that in your capture data there are instances of the hand in the same position as in waving, but not actually waving. Then do not tag those instances as positive examples of waving.

我的2cp