GuruSwiftSDK

A Swift SDK for interacting with the Guru API.

Getting Started

After adding the package as a dependency to your project, you will want to implement a controller like the following. It is a simple controller that starts inference (in response to some button click) and shows each captured frame to the user, in addition to rendering some of the inference result. Each section is described in more detail below.

import UIKit
import AVFoundation
import GuruSwiftSDK

class InferenceViewController: UIViewController {
  
  var inference: LocalVideoInference?
  @IBOutlet weak var imageView: UIImageView!
  var userLastFacing: UserFacing = UserFacing.other

  @IBAction func beingCapture(_ sender: AnyObject) {
    do {
        inference = try LocalVideoInference(
        consumer: self,
        cameraPosition: .front,
        source: "your-company-name",
        apiKey: "your-api-key"
      )
      
      Task {
        let videoId = try await inference!.start(activity: Activity.shoulder_flexion)
        print("Guru videoId is \(videoId)")
      }
    }
    catch {
      print("Unexpected error starting inference: \(error)")
    }
  }
  
  override func viewWillDisappear(_ animated: Bool) {
    Task {
      try! await inference?.stop()
    }
  }
}

extension InferenceViewController: InferenceConsumer {
  
  func consumeAnalysis(analysis: Analysis) {
    // TODO: Implement this function.
  }
  
  func consumeFrame(frame: UIImage, inference: FrameInference?) {
    if (inference != nil) {
      let painter = InferencePainter(frame: frame, inference: inference!)
        .paintLandmarkConnector(from: InferenceLandmark.leftShoulder, to: InferenceLandmark.leftElbow)
        .paintLandmarkConnector(from: InferenceLandmark.leftElbow, to: InferenceLandmark.leftWrist)
        .paintLandmarkConnector(from: InferenceLandmark.leftShoulder, to: InferenceLandmark.leftHip)
        .paintLandmarkConnector(from: InferenceLandmark.leftHip, to: InferenceLandmark.leftKnee)
        .paintLandmarkConnector(from: InferenceLandmark.leftKnee, to: InferenceLandmark.leftAnkle)
        .paintLandmarkConnector(from: InferenceLandmark.rightShoulder, to: InferenceLandmark.rightElbow)
        .paintLandmarkConnector(from: InferenceLandmark.rightElbow, to: InferenceLandmark.rightWrist)
        .paintLandmarkConnector(from: InferenceLandmark.rightShoulder, to: InferenceLandmark.rightHip)
        .paintLandmarkConnector(from: InferenceLandmark.rightHip, to: InferenceLandmark.rightKnee)
        .paintLandmarkConnector(from: InferenceLandmark.rightKnee, to: InferenceLandmark.rightAnkle)
      
      let userFacing = inference!.userFacing()
      if (userFacing != UserFacing.other) {
        userLastFacing = userFacing
      }
      if (userLastFacing == UserFacing.left) {
        painter.paintLandmarkAngle(center: InferenceLandmark.rightShoulder, from: InferenceLandmark.rightHip, to: InferenceLandmark.rightElbow, clockwise: true)
      }
      else if (userLastFacing == UserFacing.right) {
        painter.paintLandmarkAngle(center: InferenceLandmark.leftShoulder, from: InferenceLandmark.leftHip, to: InferenceLandmark.leftElbow, clockwise: false)
      }
      
      imageView.image = painter.finish()
    }
    else {
      imageView.image = frame
    }
  }
}

The member variables of the controller are:

var inference: LocalVideoInference?

The LocalVideoInference is the main engine for interacting with the GuruSwiftSDK. You will use it to start and stop the inference.

@IBOutlet weak var imageView: UIImageView!

A handle to a UIImageView that we'll use to display each captured frame.

var userLastFacing: UserFacing = UserFacing.other

A variable to store the direction the user was facing in the previous frame. We'll use this below to help us in cases where the inference confidence is low.

The beingCapture method would be called in response to the user clicking a button to start capturing.

inference = try LocalVideoInference(
  consumer: self,
  cameraPosition: .front,
  source: "your-company-name",
  apiKey: "your-api-key"
)

Task {
  let videoId = try await inference!.start(activity: Activity.shoulder_flexion)
  print("Guru videoId is \(videoId)")
}

It takes the source and apiKey, that will have been provided to you by Guru. You also specify which phone camera to use. The consumer is a reference to the object that will be called as inference is performed. It must implement the InferenceConsumer protocol. The call to start will open the camera and begin making callbacks to the consumer.

The viewWillDisappear method is called when the user navigates away. It ensures that the video capturing stops using try! await inference?.stop().

The InferenceConsumer implementation has 2 important methods: func consumeFrame(frame: UIImage, inference: FrameInference?) will be called for each frame captured. It will include the frame, which is the raw image itself, and the inference, which is the information that has been analysed for the frame. You can combine the two to draw additional information on the screen about what has been captured. In the example above, it is drawing some of the keypoints to create a skeleton and the angle between the hip, shoulder, and elbow. See the method documentation in InferencePainter for more detail on each method.

The func consumeAnalysis(analysis: Analysis) callback is invoked less frequently, and contains meta analysis about each of the frames seen so far. In here you can find information about reps that have been counted, and any extra information about those reps.

Options

Following is a list of the configurable options for LocalVideoInference:

• maxDuration: The maximum amount of time, in seconds, that recording can run for. After this amount of time video capturing will automatically terminate (and the final analysis results sent to the callback). Default is 1 minute. Note that the longer a capture runs for, the longer the delay experienced in receiving new analysis results.
• analysisPerSecond: The maximum number of frames per second to send to the server for rep counting and analysis. Default is 8 per second. Lower values will results in lower bandwidth usage, at the expense of less accurate rep counting and analysis. Higher values will use more bandwidth, but give more accurate results. Generally speaking, the faster a movement is, the higher this value should be. Note that the video is recorded at 30 fps and so setting any value higher than this will have no affect.
• recordTo: If provided, the path to a file where the video will be recorded (in addition to also being streamed to the callback).

Recording

If you wish to record the captured video then you can use an AVAssetWriter to output each captured frame to a file. See here for an example implementation.

If the video is recorded, you may choose to call uploadVideo after recording has been stopped. By uploading your video to the Guru servers, overlay videos will be built for the video. These overlay videos include rep counts and wireframes drawn over the top of the video. The returned UploadResult will contain URLs from which the overlays can be downloaded.

Requirements

This SDK requires iOS 15 or higher to function. It will throw a runtime error if run on iOS >= 13 and < 15.

It has been tested for performance on iPhone 12 and higher. iPhone 11 will function, albeit with slower performance.

Development

How to build against OpenCV

Note: these instructions are for Guru developers only.

Use the build-xcframework.sh script to package OpenCV (used by libgurucv). Its output will instruct you on how to store the artifact in S3 and update the checksum in Package.swift.

cd thirdparty
./build-xcframework.sh

How to run tests

The easiest way is to run them from the Test navigator in XCode.