华为机器学习（ML Kit）提供手部关键点辨认服务，可用于手语辨认。手部关键点辨认服务能辨认手部21个关键点，通过每个手指的方向和手语规定作比拟去找手语字母表。
利用场景
手语通常被听力和书面语有阻碍的人来应用，是收集手势蕴含日常互动中所应用的动作和手势。应用ML Kit 能够建设一个智能手语字母表识别器，它能够像一个辅助器一样将手势翻译成单词或者句子，也能够将单词或者句子翻译成手势。这里尝试的是手势当中的美国手语字母表，是基于关节，手指和手段的地位进行分类。接下来小编将会尝试从手势中收集单词“HELLO”。
开发步骤
1. 开发筹备
具体的筹备步骤能够参考华为开发者联盟，这里列举要害的开发步骤。
1.1 启动ML Kit
在华为开发者AppGallery Connect, 抉择Develop > Manage APIs。确保ML Kit 激活。
1.2 我的项目级gradle里配置Maven仓地址
buildscript { repositories { ... maven {url 'https://developer.huawei.com/repo/'} } } dependencies { ... classpath 'com.huawei.agconnect:agcp:1.3.1.301' } allprojects { repositories { ... maven {url 'https://developer.huawei.com/repo/'} } }1.3 集成SDK后，在文件头增加配置
apply plugin: 'com.android.application' apply plugin: 'com.huawei.agconnect' dependencies{ // Import the base SDK. implementation 'com.huawei.hms:ml-computer-vision-handkeypoint:2.0.2.300' // Import the hand keypoint detection model package. implementation 'com.huawei.hms:ml-computer-vision-handkeypoint-model:2.0.2.300' }1.4 将以下语句增加到AndroidManifest.xml文件中
<meta-data android:name="com.huawei.hms.ml.DEPENDENCY" android:value= "handkeypoint"/>1.5 申请摄像头权限和本地文件读取权限
<!--Camera permission--> <uses-permission android:name="android.permission.CAMERA" /> <!--Read permission--> <uses-permission android:name="android.permission.READ_EXTERNAL_STORAGE" />2. 代码开发
2.1 创立用于相机预览的Surface View，创立用于后果的Surface View。
目前咱们只在UI中显示后果，您也能够应用TTS辨认扩大和读取后果。
mSurfaceHolderCamera.addCallback(surfaceHolderCallback) private val surfaceHolderCallback = object : SurfaceHolder.Callback { override fun surfaceCreated(holder: SurfaceHolder) { createAnalyzer() } override fun surfaceChanged(holder: SurfaceHolder, format: Int, width: Int, height: Int) { prepareLensEngine(width, height) mLensEngine.run(holder) } override fun surfaceDestroyed(holder: SurfaceHolder) { mLensEngine.release() } }2.2 创立手部关键点分析器
//Creates MLKeyPointAnalyzer with MLHandKeypointAnalyzerSetting.val settings = MLHandKeypointAnalyzerSetting.Factory() .setSceneType(MLHandKeypointAnalyzerSetting.TYPE_ALL) .setMaxHandResults(2) .create()// Set the maximum number of hand regions that can be detected within an image. A maximum of 10 hand regions can be detected by default mAnalyzer = MLHandKeypointAnalyzerFactory.getInstance().getHandKeypointAnalyzer(settings)mAnalyzer.setTransactor(mHandKeyPointTransactor)2.3 开发者创立辨认后果解决类“HandKeypointTransactor”
该类MLAnalyzer.MLTransactor<T>接口，应用此类中的“transactResult”办法获取检测后果并实现具体业务。
class HandKeyPointTransactor(surfaceHolder: SurfaceHolder? = null): MLAnalyzer.MLTransactor<MLHandKeypoints> { override fun transactResult(result: MLAnalyzer.Result<MLHandKeypoints>?) { var foundCharacter = findTheCharacterResult(result) if (foundCharacter.isNotEmpty() && !foundCharacter.equals(lastCharacter)) { lastCharacter = foundCharacter displayText.append(lastCharacter) } canvas.drawText(displayText.toString(), paddingleft, paddingRight, Paint().also { it.style = Paint.Style.FILL it.color = Color.YELLOW }) }2.4 创立LensEngine
LensEngine lensEngine = new LensEngine.Creator(getApplicationContext(), analyzer)setLensType(LensEngine.BACK_LENS)applyDisplayDimension(width, height) // adjust width and height depending on the orientationapplyFps(5f)enableAutomaticFocus(true)create();2.5 运行LensEngine
private val surfaceHolderCallback = object : SurfaceHolder.Callback { // run the LensEngine in surfaceChanged() override fun surfaceChanged(holder: SurfaceHolder, format: Int, width: Int, height: Int) { createLensEngine(width, height) mLensEngine.run(holder)} }2.6 进行分析器，开释检测资源
fun stopAnalyzer() { mAnalyzer.stop() }2.7 解决 transactResult() 以检测字符
您能够应用HandKeypointTransactor类中的transtresult办法来获取检测后果并实现特定的服务。检测后果除了手部各关键点的坐标信息外，还包含手掌和每个关键点的相信值。手掌和手部关键点辨认谬误能够依据相信值过滤掉。在理论利用中，能够依据误认容忍度灵便设置阈值。
2.7.1 找到手指的方向
让咱们先假如可能手指的矢量斜率别离在X轴和Y轴上。
private const val X_COORDINATE = 0private const val Y_COORDINATE = 1假如咱们有手指别离在5个矢量上，任意手指的方向在任意工夫能够被分类为上，下，下-上，上-下，不动。
enum class FingerDirection { VECTOR_UP, VECTOR_DOWN, VECTOR_UP_DOWN, VECTOR_DOWN_UP, VECTOR_UNDEFINED} enum class Finger { THUMB, FIRST_FINGER, MIDDLE_FINGER, RING_FINGER, LITTLE_FINGER}首先将对应的关键点从后果中拆散到不同手指的关键点数组，像这样：
var firstFinger = arrayListOf<MLHandKeypoint>()var middleFinger = arrayListOf<MLHandKeypoint>()var ringFinger = arrayListOf<MLHandKeypoint>()var littleFinger = arrayListOf<MLHandKeypoint>()var thumb = arrayListOf<MLHandKeypoint>()手指上的每个关键点都对应手指的关节，通过计算关节与手指的均匀地位值之间的间隔就能够计算出斜率。依据左近关键点的坐标，查问该关键点的坐标。
拿字母H的两个简略关键点来说：
int[] datapointSampleH1 = {623, 497, 377, 312, 348, 234, 162, 90, 377, 204, 126, 54, 383, 306, 413, 491, 455, 348, 419, 521 };int [] datapointSampleH2 = {595, 463, 374, 343, 368, 223, 147, 78, 381, 217, 110, 40, 412, 311, 444, 526, 450, 406, 488, 532};用手指坐标的平均值来计算矢量
//For ForeFinger - 623, 497, 377, 312 double avgFingerPosition = (datapoints[0].getX()+datapoints[1].getX()+datapoints[2].getX()+datapoints[3].getX())/4;// find the average and subract it from the value of xdouble diff = datapointSampleH1 [position] .getX() - avgFingerPosition ;//vector either positive or negative representing the directionint vector = (int)((diff *100)/avgFingerPosition ) ;矢量的后果将会是正值或者负值，如果它是正值它会呈现X轴的正四方向，如果相同它就是负值。用这个形式对所有字母进行矢量映射，一旦你把握了所有的矢量咱们就能够用它们来进行编程。
用上述矢量方向，咱们能够分类矢量，定义第一个为手指方向枚举
private fun getSlope(keyPoints: MutableList<MLHandKeypoint>, coordinate: Int): FingerDirection { when (coordinate) { X_COORDINATE -> { if (keyPoints[0].pointX > keyPoints[3].pointX && keyPoints[0].pointX > keyPoints[2].pointX) return FingerDirection.VECTOR_DOWN if (keyPoints[0].pointX > keyPoints[1].pointX && keyPoints[3].pointX > keyPoints[2].pointX) return FingerDirection.VECTOR_DOWN_UP if (keyPoints[0].pointX < keyPoints[1].pointX && keyPoints[3].pointX < keyPoints[2].pointX) return FingerDirection.VECTOR_UP_DOWN if (keyPoints[0].pointX < keyPoints[3].pointX && keyPoints[0].pointX < keyPoints[2].pointX) return FingerDirection.VECTOR_UP } Y_COORDINATE -> { if (keyPoints[0].pointY > keyPoints[1].pointY && keyPoints[2].pointY > keyPoints[1].pointY && keyPoints[3].pointY > keyPoints[2].pointY) return FingerDirection.VECTOR_UP_DOWN if (keyPoints[0].pointY > keyPoints[3].pointY && keyPoints[0].pointY > keyPoints[2].pointY) return FingerDirection.VECTOR_UP if (keyPoints[0].pointY < keyPoints[1].pointY && keyPoints[3].pointY < keyPoints[2].pointY) return FingerDirection.VECTOR_DOWN_UP if (keyPoints[0].pointY < keyPoints[3].pointY && keyPoints[0].pointY < keyPoints[2].pointY) return FingerDirection.VECTOR_DOWN } }return FingerDirection.VECTOR_UNDEFINED获取每个手指的方向并且贮存在一个数组里。
xDirections[Finger.FIRST_FINGER] = getSlope(firstFinger, X_COORDINATE)yDirections[Finger.FIRST_FINGER] = getSlope(firstFinger, Y_COORDINATE ）2.7.2 从手指方向找到字符：
当初咱们把它当作惟一的单词“HELLO”，它须要字母H,E,L,O。它们对应的X轴和Y轴的矢量如图所示。
假如：手的方向总是竖向的。让手掌和手段与手机平行，也就是与X轴成90度。姿态至多放弃3秒用来记录字符。
开始用字符映射矢量来查找字符串
// Alphabet Hif (xDirections[Finger.LITTLE_FINGER] == FingerDirection.VECTOR_DOWN_UP && xDirections [Finger.RING_FINGER] == FingerDirection.VECTOR_DOWN_UP && xDirections [Finger.MIDDLE_FINGER] == FingerDirection.VECTOR_DOWN && xDirections [Finger.FIRST_FINGER] == FingerDirection.VECTOR_DOWN && xDirections [Finger.THUMB] == FingerDirection.VECTOR_DOWN) return "H" //Alphabet Eif (yDirections[Finger.LITTLE_FINGER] == FingerDirection.VECTOR_UP_DOWN && yDirections [Finger.RING_FINGER] == FingerDirection.VECTOR_UP_DOWN && yDirections [Finger.MIDDLE_FINGER] == FingerDirection.VECTOR_UP_DOWN && yDirections [Finger.FIRST_FINGER] == FingerDirection.VECTOR_UP_DOWN && xDirections [Finger.THUMB] == FingerDirection.VECTOR_DOWN) return "E" if (yDirections[Finger.LITTLE_FINGER] == FingerDirection.VECTOR_UP_DOWN && yDirections [Finger.RING_FINGER] == FingerDirection.VECTOR_UP_DOWN && yDirections [Finger.MIDDLE_FINGER] == FingerDirection.VECTOR_UP_DOWN && yDirections [Finger.FIRST_FINGER] == FingerDirection.VECTOR_UP && yDirections [Finger.THUMB] == FingerDirection.VECTOR_UP) return "L" if (xDirections[Finger.LITTLE_FINGER] == FingerDirection.VECTOR_UP && xDirections [Finger.RING_FINGER] == FingerDirection.VECTOR_UP && yDirections [Finger.THUMB] == FingerDirection.VECTOR_UP)return "O"3. 画面和后果
4．更多技巧和窍门
1. 当扩大到26个字母时，误差很更多。为了更精准的扫描须要2-3秒，从2-3秒的工夫寻找和计算最有可能的字符，这能够缩小字母表的误差。
2. 为了能反对所有方向，在X-Y轴上减少8个或者更多的方向。首先，需要求出手指的度数和对应的手指矢量。
总结
这个尝试是强力坐标技术，它能够在生成矢量映射后扩大到所有26个字母，方向也能够扩大所有8个方向，所以它会有26*8*5个手指=1040个矢量。为了更好的解决这一问题，咱们能够利用手指的一阶导数函数来代替矢量从而简化计算。
咱们能够加强其它的去代替创立矢量，能够应用图像分类和训练模型，而后应用自定义模型。这个训练是为了查看华为ML Kit应用关键点解决个性的可行性。
理解更多相干内容>>
拜访华为机器学习服务官网
获取华为机器学习服务开发领导文档
华为HMS Core官方论坛
华为机器学习开源仓地址：GitHub、Gitee
解决集成问题请到Stack Overflow
点击关注，第一工夫理解HMS Core最新技术~