Sprint 5 - User Stories
Sprint 5 is divided in three task groups: conceptual, implementation, and off-scope.
Conceptual Tasks
Distinguish in Detection and Tracking.Get a book on Object Tracking in Videos with OpenCV oder dlib; reiterate Howse 2020 book.- Think of how to model: Consider orientation of leaves coming from stem.
- Think of how to model: Predominant response of individuals in larger groups is to maintain spacing with near neighbors, decelerating or accelerating to avoid those very close behind or ahead, respectively, or to turn away from neighbors who approach very closely from either side.
Feature extraction of GAI using Gabor filters: is filtering operation of the given GEI with the Gabor filter of size u and orientation v. > Model stem?Gait feature dimension reduction based on Kernel Fisher Analysis: Idea is to yield nonlinear discriminant analysis in the higher space as F (Youtube Video).- Check what is ‘butterworth filter’ as used in Nature article for the interpolation of the bird flights.
Read over Human Gait papers.
Implementation Tasks
Detection
Standardize Input Video > Conversion to 640x480p Video. > For high / low values recognition.Contrast Thresholding: Try Thresholding Colour Images from Mastering OpenCV4 (internal screenshot: green tresh)Integrate a width, height threshold > parameter has to be greater / smaller then w / h value when being tracked.- Try thresholding techniques of Scikit-Image: Niblack and Sauvola are local thresholding techniques (should be better for plants because background is not uniform).
- Try FindContours():
Image and extreme points: extreme_points calculates the four extreme points defining a given contour.- cv2.approxPolyDP(): returns a contour approximation of given contour based on the given precision (using Douglas Peucker)
- Check Hu Moment Invariants
- Feature Detection mechanisms with OpenCV are Harris Corner Detection, Shi-Tomasi, SIFT, SURF< FAST, BRIEF< ORB. Try them.
Check histogram for your flying paprika.
Tracking
- Feature Detection mechanisms with OpenCV are Harris Corner Detection, Shi-Tomasi, SIFT, SURF< FAST, BRIEF< ORB. Try them.
- Motion prediction: Fish motion is a four-dim vector (x,y, v(x), v(y)) i.e. x, y value and velocity of both moving. Then, movement is predicted by Kalman filter. Inclusion of a compensation window (as only movements ±45 degrees are reasonable).
- Feature matching: Active contour model is used for the extraction of fish head contours. Matching cascade: width matching, area matching and gray matching.
Get a book on Object Tracking in Videos with OpenCV oder dlib- Trajectory linking: Use of state variables for linking. »> Figure out how they did this.
Off-Scope Tasks
Data Processing & Plotting
- Extract 5/10/20/30 tracking points max > use a dropdown in beginning; adapt graphs to it > max 5 lines per graph
- Own graph for MFCC; 5 tracking points per graph; draw lines into image.
Happiness Indicator
- Height is max value. Low is min value. Height region is max - 15%; Low Region is min + 15%.
Weekend Projects
- Do the FLASK web project on Face Recognition & host on your web page as a weekend project.
- Change blog style to make articles more readable.