從iOS 14和macOS 11開始，Vision增加了識別人體姿勢的強(qiáng)大功能。他可以識別人體的19個關(guān)鍵點(diǎn)。如圖所示：

人體關(guān)鍵點(diǎn).png

實(shí)現(xiàn)

1.發(fā)起一個請求

使用Vision框架，通過VNDetectHumanBodyPoseRequest提供身體姿勢檢測功能。 下面代碼演示了，如何從CGImage中檢測身體關(guān)鍵點(diǎn)。

// Get the CGImage on which to perform requests.
guard let cgImage = UIImage(named: "bodypose")?.cgImage else { return }

// Create a new image-request handler.
let requestHandler = VNImageRequestHandler(cgImage: cgImage)

// Create a new request to recognize a human body pose.
let request = VNDetectHumanBodyPoseRequest(completionHandler: bodyPoseHandler)

do {
    // Perform the body pose-detection request.
    try requestHandler.perform([request])
} catch {
    print("Unable to perform the request: \(error).")
}

2.處理結(jié)果

請求處理完之后，會調(diào)用完成的閉包，通過閉包，可以獲取到檢測結(jié)果和錯誤信息。 如果正常檢測到人體關(guān)鍵點(diǎn)，將以VNHumanBodyPoseObservation數(shù)組的形式返回。VNHumanBodyPoseObservation中包含識別到的關(guān)鍵點(diǎn)和一個置信度分?jǐn)?shù)，置信度越大，說明識別的精度越高。

func bodyPoseHandler(request: VNRequest, error: Error?) {
    guard let observations =
            request.results as? [VNHumanBodyPoseObservation] else { 
        return 
    }
    
    // Process each observation to find the recognized body pose points.
    observations.forEach { processObservation($0) }
}

3.獲取關(guān)鍵點(diǎn)

可以通過VNHumanBodyPoseObservation.JointName來獲取對應(yīng)的關(guān)鍵點(diǎn)的坐標(biāo)。注意，recognizedPoints(_:) 方法返回的點(diǎn)取值范圍[0, 1]，原點(diǎn)位于左下角，實(shí)際使用中需要進(jìn)行轉(zhuǎn)換。

func processObservation(_ observation: VNHumanBodyPoseObservation) {
    
    // Retrieve all torso points.
    guard let recognizedPoints =
            try? observation.recognizedPoints(.torso) else { return }
    
    // Torso joint names in a clockwise ordering.
    let torsoJointNames: [VNHumanBodyPoseObservation.JointName] = [
        .neck,
        .rightShoulder,
        .rightHip,
        .root,
        .leftHip,
        .leftShoulder
    ]
    
    // Retrieve the CGPoints containing the normalized X and Y coordinates.
    let imagePoints: [CGPoint] = torsoJointNames.compactMap {
        guard let point = recognizedPoints[$0], point.confidence > 0 else { return nil }
        
        // Translate the point from normalized-coordinates to image coordinates.
        return VNImagePointForNormalizedPoint(point.location,
                                              Int(imageSize.width),
                                              Int(imageSize.height))
    }
    
    // Draw the points onscreen.
    draw(points: imagePoints)
}

拓展

除了使用 Vision 的 VNDetectHumanBodyPoseRequest 來做人體姿勢識別之外，還可以使用 CoreML 來實(shí)現(xiàn)。官方示例：Detecting Human Body Poses in an Image ，此示例可以運(yùn)行在 iOS 13及以上版本。