Master in Computer Vision, Barcelona (2017-2018) - M6 Video Analysis
We are Team 7:
Roger Marí. Email: roger.mari01@estudiant.upf.edu
Joan Sintes. Email: joan.sintes01@estudiant.upf.edu
Àlex Palomo. Email: alex.palomo01@estudiant.upf.edu
Àlex Vicente. Email: alex.vicente01@estudiant.upf.edu
This 5-week project presents a series of experiments related to video analysis for traffic monitoring.
We will employ basic concepts and techniques related to video sequences mainly for surveillance applications,
divided in 4 main stages: background estimation, foreground segmentation, video stabilization and region tracking.
Check our slides for the 5th Workshop on “Road Traffic Monitoring” here.
The report on our traffic monitoring system can be found here.
Task 1. Background substraction. Segmentation metrics. Precision and recall.
Task 2. Background susbtraction. Segmentation metrics. Temporal analysis.
Task 3. Optical flow evaluation metrics. Mean Squared Error and Percentage of Erroneous Pixels in Non-occluded areas.
Task 4. Background substraction. Evaluation of de-synchornized results.
Task 5. Visual representation of optical flow.
- Check the information about how to run the code here
Task 1. Gaussian modelling. Evaluation by means of F-score vs alpha and AUC (Precision-Recall curves).
Task 2. Adaptive modelling. Comparison between adaptive and non-adaptive methods via F-score and AUC.
Task 3. Comparison with state of the art. Methods from Tasks 1 and 2 are compared to MOG and MOG2.
Task 4. Gaussian modelling taking into account color. RGB and YCbCr colorspaces used.
- Check the information about how to run the code here
Task 1. Hole filling to complete objects in the foreground.
Task 2. Area filtering to remove noise from the background.
Task 3. Morphological operators (closing + hole filling) to boost perfromance.
Task 4. Shadow detection and removal (pixel based methods using the HSV colorspace).
Task 5. Improvement in Precision-Recall curves with respect to the best configuration from week 2.
- Check the information about how to run the code here
Task 1.1. Optical Flow with Block Matching (using MSE as matching cost).
Task 1.2. Block Matching vs other techniques (Farnebäck's method and TV-L1 Optical Flow).
Task 2.1. Video Stabilization with Block Matching (experiments with 2 approaches).
Task 2.2. Block Matching Video Stabilization vs other techiques (Pyramidal Lucas-Kanade and Homography-based).
Task 2.3. Video Stabilization of videos of our own.
- Check the information about how to run the code here
Task 1.1. Tracking with Kalman Filter (and bounding box merging).
Task 1.2. Tracking with other methods. The Median-Flow tracker.
Task 2. Speed estimation via homography rectification.
Task 3. Own study: car density (cars/frame), traffic rate (cars/minute) and infraction detection (speed limit 80km/h).
- Check the information about how to run the code here