iOS MachineLearning 系列（5）—— 视频中的物体运动跟踪

lazy var handler = VNSequenceRequestHandler()

lazy var request: VNTrackRectangleRequest = {    
    let req = VNTrackRectangleRequest(rectangleObservation: observation) { result, error in
        // 处理结果
        if let error {
            print(error)
        }
        // 处理结果
        self.handleResult(request: result as! VNTrackRectangleRequest)
    }
    // 选择快速模式
    req.trackingLevel = .fast
    return req
}()

// 预检测得到的
var observation = VNRectangleObservation(boundingBox: CGRect(x: 0.3728713095188141, y: 0.833836019039154, width: 0.16493645310401917, height: 0.07572066783905029))

func readVideo() {
    // 视频路径
    let videoURL = URL(fileURLWithPath: Bundle.main.path(forResource: "video1", ofType: ".mp4")!)
    // 读取视频资源
    let videoAsset = AVURLAsset(url: videoURL)
    // 创建视频资源解析器
    let videoProcessor = AVAssetImageGenerator(asset: videoAsset)
    videoProcessor.requestedTimeToleranceBefore = CMTime.zero
    videoProcessor.requestedTimeToleranceAfter = CMTime.zero
    // 获取视频时长
    let durationSeconds: Float64 = CMTimeGetSeconds(videoAsset.duration)
    // 存储要截取的视频帧时间点
    var times = [NSValue]()
    // 以每秒60帧为标准，获取总帧数
    let totalFrames: Float64 = durationSeconds * 60
    // 定义 CMTime 即请求缩略图的时间间隔
    for i in 0...Int(totalFrames) {
       let timeFrame = CMTimeMake(value: Int64(i), timescale: 60)
       let timeValue = NSValue(time: timeFrame)
       times.append(timeValue)
    }
    // 进行图片解析
    videoProcessor.generateCGImagesAsynchronously(forTimes: times) { time, cgImage, actualTime, resultCode, error  in
        if let cgImage = cgImage {
            let image = UIImage(cgImage: cgImage)
            self.images.append(image)
        }
    }
}

func start() {
    var count = 0
    // 这里定时器的作用是逐帧的刷新页面，同时进行追踪
    Timer.scheduledTimer(withTimeInterval: 0.03, repeats: true) { t in
        if count < self.images.count {
            // 设置页面展示的图片
            self.imageView.image = self.images[count]
            // 将inputObservation设置为上一次的分析结果
            self.request.inputObservation = self.observation
            // 进行追踪分析
            try? self.handler.perform([self.request], on: self.images[count].cgImage!, orientation: .up)
            count += 1
        } else {
            // 当循环结束时，设置isLastFrame表情请求已经到了最后一帧
            self.request.isLastFrame = true
            // 停止定时器
            t.invalidate()
            print("end")
        }
    }
    print(images.count)
}

func handleResult(request: VNTrackRectangleRequest) {
    print(request.results)
    for r in request.results ?? [] {
        guard let result = r as? VNRectangleObservation else {
            return
        }
        observation = result
        var box = result.boundingBox
        // 坐标系转换
        box.origin.y = 1 - box.origin.y - box.size.height
        print("box:", result.boundingBox)
        DispatchQueue.main.async {
            let size = self.imageView.frame.size
            self.boxView.frame = CGRect(x: box.origin.x * size.width, y: box.origin.y * size.height, width: box.size.width * size.width, height: box.size.height * size.height)
        }
    }
}

open class VNTrackingRequest : VNImageBasedRequest {
    // 输入追踪区域的Observation对象 每次根据中要刷新
    open var inputObservation: VNDetectedObjectObservation
    // 追踪模式
    open var trackingLevel: VNRequestTrackingLevel
    // 是否是最后一帧，如果设置为true，将停止后续分析
    open var isLastFrame: Bool
}

public enum VNRequestTrackingLevel : UInt, @unchecked Sendable {
    // 精准优先
    case accurate = 0
    // 速度优先
    case fast = 1
}