检测侧面和后面的骷髅 [英] Detecting Sideways and Backwards Skeletons

问题描述

Kinect SDK v1-1.8的一个问题是当用户没有面对Kinect时骨架检测不准确。

站在侧面时（Kinect正好在你的身上）向右或向左）你会得到非常紧张的骨骼关节。 （我知道这是因为遮挡而发生的，但我相信必须有一种方法可以获得平稳的预测）

当倒立（Kinect落后于你）时，它会像你一样发现你面对他Kinect并切换左右关节（例如：ShoulderLeft实际上是你的右肩）。

One of the problems with Kinect SDK v1-1.8 is the inaccuracies in Skeletal Detection when the user isn't facing the Kinect.
When standing sideways (the Kinect is exactly on your right or left) you would get very jittery skeleton joints. (I understand that this happens due to occlusion, but I believe there must be a way to get a smooth prediction)
When standing backwards (the Kinect is behind you) it would detect you as if you were facing he Kinect and switching the left and right joints (eg: ShoulderLeft would actually be your right shoulder).

我的问题是，Kinect SDK v2修复了这些问题吗？在多大程度上？我看到骷髅现在更"解剖学正确"了。并且总体抖动较小，所以我希望这些问题至少部分修复，但我想确定。

My question is, are these problems fixed with Kinect SDK v2? to what extent?
I've seen that the skeletons are now more "anatomically correct" and have less overall jitter, so I expect these problem to be at least partially fixed, but I want to be sure.

谢谢：）

推荐答案

我们不会为不面向传感器的人员进行身体跟踪训练，因此数据可能不符合您的要求。看看身体跟踪演示并用不同的形状替换正确的关节以查看行为。作为新SDK的一部分，除了增加5个额外关节外，我们还有
解剖学校正位置。请参阅第9频道发布的Jumpstart视频系列：
http://aka.ms/k4wv2jumpstart 这将在第一部分进一步讨论在模块7中结合视频和一些更高级的跟踪主题。

We don't train the body tracking for people not facing the sensor so the data may not be exact to what you want. Have a look at the body tracking demo and replace the right joints with a different shape to see the behavior. As part of the new SDK we have anatomically corrected positions in addition to added 5 additional joints. See the Jumpstart video series posted on Channel 9: http://aka.ms/k4wv2jumpstart this is discussed further in the first couple videos and some more advanced tracking topics in module 7.

身体跟踪旨在仅跟踪面向传感器的那些，因此它不应该检测最初面向的人。锁定只有在能够检测到人的头部和肩膀时才会发生。当你拥有它时，除了臀部/肩膀之间的矢量之外，你还可以使用面部
跟踪来跟踪头部，以确定被跟踪人员的平面，并确定它们如果开始朝向它们时如何旋转。通常情况下，很多关节都会在某一点被推断出来。

Body tracking is designed to only track those that are facing the sensor, so it should not detect someone facing away initially. Lock-on will occur only when it can detect the head and both shoulders of the person. When you have that you can use face tracking to track the head in addition to the the vector between the hips/shoulders to determine the plane of the tracked person and determine how they rotate if they are starting to face away. Typically a lot of the joints will become inferred at one point.

这篇关于检测侧面和后面的骷髅的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！