Motion detectorπŸ”—


This recipe shows how to make an app that detects motion by monitoring changes in successive input images and raises an alarm if an object of the specified type (e.g.Β a person, a dog or a car) is seen in the image. Furthermore, the image which triggered the alarm can be saved to a file.

Detailed descriptionπŸ”—

The big picture of the motion detector app is given below. The processing graph is too large to fit on one page, which is why we’ll walk it through in smaller pieces.

Big picture of the motion detector

Big picture of the motion detectorπŸ”—

  • The images originate from the default webcam as configured in Image Source. Notice that the default camera is selected automatically by setting the Camera Id as /^webcam/, a regular expression that finds the first webcam connected to the computer.

  • Convert Colors converts possibly encoded or compressed images to RGB format. This block is actually not needed as long as the images come from Image Source which decodes the images automatically. However, Convert Colors serves as a convenient entry point in case you later want to use the app as an API to which a client can send compressed images for analysis.

  • Detect Motion compares the input image to the background image which is a weighted average of all previous images. If more than 1% of the pixels (i.e.Β Motion Threshold = 0.01) differ from the background by more than 10 intensity levels (i.e.Β Threshold = 10), it declares that motion has been detected by setting output Motion Detected as true. Parameters Alpha1 and Alpha2 control how fast the background image adapts to changes in input images.

  • Gate passes through the input image and Motion Detected flag only if motion has been detected-

Gate passes through only images in which motion has been detected.

Gate passes through only images in which motion has been detected.πŸ”—

  • YOLO ONNX model tries to recognize objects in the image. The processing path from Scale Image to Process YOLO Result is identical to the YOLO Classifier recipe. The output is a list of bounding boxes of found objects as pairs of Frame and Size matrices, and a corresponding list of object classes as integers between 0 and 19.

YOLO recognizes an object in the image.

YOLO recognizes an object in the image.πŸ”—

  • Information about detected motion, classes of objects which possibly caused the motion, as well as locations of the objects are passed to JavaScript which makes the final decision on whether to raise an alarm. Inputs, outputs and the code for the script are given below.

    YOLO recognizes 20 object classes. The name of the class can be configured using the Object Type parameter. The class names are listed in the script below. If the Object Type parameter is left empty, the alarm is raised whenever any motion is detected.

A script decides if an alarm should be raised.

A script decides if an alarm should be raised.πŸ”—

// Default values for frame & size
var outFrame = VisionAppster.identity(4);
var outSize = VisionAppster.doubleMatrix(1,2);
outSize.setEntry(0, 0, 0);
outSize.setEntry(0, 1, 0);

var yoloClasses = ["aeroplane", "bicycle", "bird", "boat", "bottle",
                   "bus", "car", "cat", "chair", "cow",
                   "table", "dog", "horse", "motorbike", "person",
                   "pottedplant", "sheep", "sofa", "train", "tv"];
var classId = yoloClasses.indexOf($i.objectType);

var classIndex = -1;
for (var i = 0; i < $i.classIndex.rows; ++i)
  if ($i.classIndex.entry(i, 0) == classId)
      classIndex = i;

if (classIndex >= 0 && $i.motionDetected)
    outFrame = $i.frame.sub(4 * classIndex, 4);
    outSize = $i.size.sub(classIndex, 1);
    $o.objectDetected = true;
else if (classId < 0 && $i.motionDetected)
  $o.objectDetected = true;
  $o.objectDetected = false;

$o.frame = outFrame;
$o.size = outSize;

If a motion has been detected and the image contains an object of interest, ObjectDetected flag is set to true. Remember however that the alarm may have been triggered also by motion of something else than the recognized object.

In this example the frame and size outputs (i.e.Β the bounding box information) are not used. You can use them for example by dragging and dropping Crop Image tool on the canvas and connecting the output of Gate: Motion tool to Image input of Crop Image. Then connect Frame and Size outputs of the script to the corresponding input of Crop Image tool.

Object Detected flag is connected to yet another script. The script controls where the interesting images will be saved. The file path where images will be saved is given in parameter File Path. The image format will be jpg and files are named as image1.jpg, image2.jpg, image3.jpg and so on. The numbering will wrap around after Max File Count have been saved and old images will be overwritten. No images will be saved if Max File Count is set as zero.

A script controls if and where the image should be saved,

A script controls if and where the image should be saved,πŸ”—

if ($s.fileCount === undefined)
    $s.fileCount = 0;

if ($i.maxFileCount > 0 && $i.objectDetected)
    if ($s.fileCount > $i.maxFileCount)
      $s.fileCount = 1;
    $o.filePath = $i.filePath + "/image" + $s.fileCount + ".jpg";
    $o.writeEnabled = true;
    $o.filePath = "";
    $o.writeEnabled = false;

A gate gets the Write Enabled flag from the script and uses it to decide whether to pass the image and file path to Save Image.

Images of interest are saved to jpg files,

Images of interest are saved to jpg files,πŸ”—

Finally, an example of the app in action. The object of interest has been configured as a dog. The app has detected both motion and a dog, so an alarm has been raised (i.e.Β Object Detected output is True) and the image has been saved to a file.

An example of dog detection cam.

An example of dog detection cam.πŸ”—