基于人脸位置横屏转竖屏裁剪时导出会话失败问题

1. 先定位ExportSession失败的核心原因 #

添加人脸跟踪指令后导出失败，大概率是以下几个细节没处理好：

• 检查faceTrackInstructions的时间范围是否与视频片段的时间范围完全匹配，AVFoundation对时间精度要求极高，毫秒级偏差都会导致导出失败。
• 确认人脸坐标到视频渲染坐标的转换逻辑是否正确：Vision的坐标系是图像归一化坐标（原点左上，y轴向下），而AVFoundation的视频渲染坐标系会受视频方向影响，必须做翻转/平移修正。
• 避免使用固定分辨率的导出预设，优先用AVAssetExportPresetHighestQuality，动态裁剪可能导致分辨率不兼容固定预设。
• 导出前必须调用exportSession.canExport()验证，若返回false，直接打印exportSession.error?.userInfo，里面会有具体的错误原因（比如AVErrorInvalidVideoComposition通常是指令矩阵或时间范围错误）。

2. 正确实现人脸跟踪驱动的动态裁剪逻辑 #

步骤1：预分析视频帧，记录人脸轨迹 #

用Vision批量提取视频关键帧的人脸位置，保存时间戳对应的人脸 bounding box：

func getFaceTracks(from asset: AVAsset) async throws -> [CMTime: CGRect] {
    guard let videoTrack = asset.tracks(withMediaType: .video).first else {
        throw NSError(domain: "TrackError", code: -1, userInfo: [NSLocalizedDescriptionKey: "无视频轨道"])
    }
    
    let reader = try AVAssetReader(asset: asset)
    let outputSettings = [kCVPixelBufferPixelFormatTypeKey as String: kCVPixelFormatType_32BGRA]
    let trackOutput = AVAssetReaderTrackOutput(track: videoTrack, outputSettings: outputSettings)
    reader.add(trackOutput)
    reader.startReading()
    
    var faceTimestamps = [CMTime: CGRect]()
    let faceRequest = VNDetectFaceRectanglesRequest()
    let videoNaturalSize = videoTrack.naturalSize.applying(videoTrack.preferredTransform).size
    
    while let sampleBuffer = trackOutput.copyNextSampleBuffer() {
        let timestamp = CMSampleBufferGetPresentationTimeStamp(sampleBuffer)
        guard let imageBuffer = CMSampleBufferGetImageBuffer(sampleBuffer) else { continue }
        
        let handler = VNImageRequestHandler(cvPixelBuffer: imageBuffer, orientation: .up)
        try handler.perform([faceRequest])
        
        if let faceObservation = faceRequest.results?.first {
            // 转换Vision归一化坐标到视频渲染坐标
            let normalizedRect = faceObservation.boundingBox
            let faceRect = CGRect(
                x: normalizedRect.origin.x * videoNaturalSize.width,
                y: (1 - normalizedRect.origin.y - normalizedRect.height) * videoNaturalSize.height,
                width: normalizedRect.width * videoNaturalSize.width,
                height: normalizedRect.height * videoNaturalSize.height
            )
            faceTimestamps[timestamp] = faceRect
        }
    }
    return faceTimestamps
}

步骤2：构建动态裁剪的视频合成指令 #

基于人脸轨迹，为每个时间点生成裁剪变换矩阵，确保人脸始终处于竖屏画面中心：

func buildDynamicCropComposition(asset: AVAsset, faceTracks: [CMTime: CGRect]) -> AVMutableVideoComposition {
    guard let videoTrack = asset.tracks(withMediaType: .video).first else {
        return AVMutableVideoComposition()
    }
    
    let targetSize = CGSize(width: 1080, height: 1920) // 竖屏目标分辨率
    let composition = AVMutableVideoComposition()
    composition.renderSize = targetSize
    composition.frameDuration = videoTrack.minFrameDuration
    composition.renderScale = 1.0
    
    let layerInstruction = AVMutableVideoCompositionLayerInstruction(assetTrack: videoTrack)
    let mainInstruction = AVMutableVideoCompositionInstruction()
    mainInstruction.timeRange = CMTimeRange(start: .zero, duration: asset.duration)
    mainInstruction.layerInstructions = [layerInstruction]
    
    // 按时间戳排序，插入关键帧变换
    let sortedTimestamps = faceTracks.keys.sorted { $0 < $1 }
    let videoSize = videoTrack.naturalSize.applying(videoTrack.preferredTransform).size
    
    for timestamp in sortedTimestamps {
        guard let faceRect = faceTracks[timestamp] else { continue }
        
        // 计算裁剪中心：以人脸中心为基准，限制在视频边界内
        let cropCenterX = max(faceRect.midX, targetSize.width/2)
        let finalCenterX = min(cropCenterX, videoSize.width - targetSize.width/2)
        let cropCenterY = max(faceRect.midY, targetSize.height/2)
        let finalCenterY = min(cropCenterY, videoSize.height - targetSize.height/2)
        
        // 生成平移变换，将视频平移至人脸中心对齐裁剪区域中心
        let transform = CGAffineTransform(
            translationX: -(finalCenterX - targetSize.width/2),
            y: -(finalCenterY - targetSize.height/2)
        )
        layerInstruction.setTransform(transform, at: timestamp)
        
        // 为相邻时间点添加插值，保证裁剪过渡平滑
        if let nextTimestamp = sortedTimestamps.first(where: { $0 > timestamp }) {
            let timeRange = CMTimeRange(start: timestamp, duration: CMTimeSubtract(nextTimestamp, timestamp))
            layerInstruction.setTransformRamp(fromStart: transform, toEnd: transform, timeRange: timeRange)
        }
    }
    
    composition.instructions = [mainInstruction]
    return composition
}

步骤3：执行导出操作 #

func exportVideo(asset: AVAsset, composition: AVMutableVideoComposition, outputURL: URL) async throws {
    let exportSession = AVAssetExportSession(asset: asset, presetName: AVAssetExportPresetHighestQuality)!
    exportSession.videoComposition = composition
    exportSession.outputURL = outputURL
    exportSession.outputFileType = .mp4
    exportSession.shouldOptimizeForNetworkUse = true
    
    guard exportSession.canExport() else {
        throw exportSession.error ?? NSError(domain: "ExportError", code: -1, userInfo: [NSLocalizedDescriptionKey: "导出会话不可用"])
    }
    
    try await withCheckedThrowingContinuation { continuation in
        exportSession.exportAsynchronously {
            switch exportSession.status {
            case .completed:
                continuation.resume()
            case .failed, .cancelled:
                let error = exportSession.error ?? NSError(domain: "ExportError", code: -1, userInfo: [NSLocalizedDescriptionKey: "导出失败或取消"])
                continuation.resume(throwing: error)
            default:
                break
            }
        }
    }
}

3. 关键注意事项 #

• 坐标系转换：一定要注意Vision与AVFoundation的坐标系差异，尤其是视频带有旋转方向时，需要结合track.preferredTransform做修正。
• 边界限制：计算裁剪中心时必须限制在视频原始尺寸内，避免出现黑边或画面拉伸。
• 性能优化：逐帧分析人脸会拖慢速度，可改为每秒分析1-2帧，利用AVFoundation的变换插值实现平滑过渡。
• 错误排查：导出失败时一定要打印exportSession.error的完整信息，这是定位问题最快的方式。

内容的提问来源于stack exchange，提问作者John