av/filter/knn.go

// +build !circleci

/*
DESCRIPTION
  A filter that detects motion and discards frames without motion. The
  filter uses a K-Nearest Neighbours (KNN) to determine what is
  background and what is foreground.

AUTHORS
  Ella Pietraroia <ella@ausocean.org>

LICENSE
  KNN.go is Copyright (C) 2019 the Australian Ocean Lab (AusOcean)

  It is free software: you can redistribute it and/or modify them
  under the terms of the GNU General Public License as published by the
  Free Software Foundation, either version 3 of the License, or (at your
  option) any later version.

  It is distributed in the hope that it will be useful, but WITHOUT
  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
  for more details.

  You should have received a copy of the GNU General Public License
  in gpl.txt.  If not, see http://www.gnu.org/licenses.
*/

package filter

import (
	"fmt"
	"image"
	"image/color"
	"io"

	"gocv.io/x/gocv"
)

// KNN is a filter that provides basic motion detection. KNN is short for
// K-Nearest Neighbours method.
type KNN struct {
	dst     io.WriteCloser                // Destination to which motion containing frames go.
	area    float64                       // The minimum area that a contour can be found in.
	bs      *gocv.BackgroundSubtractorKNN // Uses the KNN algorithm to find the difference between the current and background frame.
	knl     gocv.Mat                      // Matrix that is used for calculations.
	debug   bool                          // If true then debug windows with the bounding boxes and difference will be shown on the screen.
	windows []*gocv.Window                // Holds debug windows.
	hold    [][]byte                      // Will hold all frames up to hf (so only every hf frame is motion detected).
	hf      int                           // The number of frames to be held.
	hfCount int                           // Counter for the hold array.
	scale   float64                       // The factor that frames will be downscaled by for motion detection.
}

// NewKNN returns a pointer to a new KNN filter struct.
func NewKNN(dst io.WriteCloser, area, threshold float64, history, kernelSize int, debug bool, hf int, scaleFactor int) *KNN {
	bs := gocv.NewBackgroundSubtractorKNNWithParams(history, threshold, false)
	k := gocv.GetStructuringElement(gocv.MorphRect, image.Pt(kernelSize, kernelSize))
	var windows []*gocv.Window
	if debug {
		windows = []*gocv.Window{gocv.NewWindow("KNN: Bounding boxes"), gocv.NewWindow("KNN: Motion")}
	}
	return &KNN{dst, area, &bs, k, debug, windows, make([][]byte, hf-1), hf, 0, 1 / float64(scaleFactor)}
}

// Implements io.Closer.
// Close frees resources used by gocv, because it has to be done manually, due to
// it using c-go.
func (m *KNN) Close() error {
	m.bs.Close()
	m.knl.Close()
	for _, window := range m.windows {
		window.Close()
	}
	return nil
}

// Implements io.Writer.
// Write applies the motion filter to the video stream. Only frames with motion
// are written to the destination encoder, frames without are discarded.
func (m *KNN) Write(f []byte) (int, error) {
	if m.hfCount < (m.hf - 1) {
		m.hold[m.hfCount] = f
		m.hfCount++
		return len(f), nil
	}
	m.hfCount = 0

	img, err := gocv.IMDecode(f, gocv.IMReadColor)
	if err != nil {
		return 0, fmt.Errorf("can't decode image: %w", err)
	}
	defer img.Close()

	imgDelta := gocv.NewMat()
	defer imgDelta.Close()

	// Downsize image to speed up calculations.
	gocv.Resize(img, &img, image.Point{}, m.scale, m.scale, gocv.InterpolationNearestNeighbor)

	// Seperate foreground and background.
	m.bs.Apply(img, &imgDelta)

	// Threshold imgDelta.
	gocv.Threshold(imgDelta, &imgDelta, 25, 255, gocv.ThresholdBinary)

	// Remove noise.
	gocv.Erode(imgDelta, &imgDelta, m.knl)
	gocv.Dilate(imgDelta, &imgDelta, m.knl)

	// Fill small holes.
	gocv.Dilate(imgDelta, &imgDelta, m.knl)
	gocv.Erode(imgDelta, &imgDelta, m.knl)

	// Find contours and reject ones with a small area.
	var contours [][]image.Point
	allContours := gocv.FindContours(imgDelta, gocv.RetrievalExternal, gocv.ChainApproxSimple)
	for _, c := range allContours {
		if gocv.ContourArea(c) > m.area {
			contours = append(contours, c)
		}
	}

	// Draw debug information.
	if m.debug {
		for _, c := range contours {
			rect := gocv.BoundingRect(c)
			gocv.Rectangle(&img, rect, color.RGBA{0, 0, 255, 0}, 1)
		}

		if len(contours) > 0 {
			gocv.PutText(&img, "Motion", image.Pt(32, 32), gocv.FontHersheyPlain, 2.0, color.RGBA{255, 0, 0, 0}, 2)
		}

		m.windows[0].IMShow(img)
		m.windows[1].IMShow(imgDelta)
		m.windows[0].WaitKey(1)
	}

	// Don't write to destination if there is no motion.
	if len(contours) == 0 {
		return len(f), nil
	}

	// Write to destination, past 4 frames then current frame.
	for i, h := range m.hold {
		_, err := m.dst.Write(h)
		m.hold[i] = nil
		if err != nil {
			return len(f), fmt.Errorf("could not write previous frames: %w", err)
		}
	}
	return m.dst.Write(f)
}