Archive for the ‘Research’ Category
T IS the end of an era as another project comes to an important milestone. I was preparing a lot of code for upload last week. I wanted to wait a while, at the very least until I had also uploaded the accompanying papers. A 2011-2012 Technical Report [HTML, PDF] about Identity Verification and “Car Navigation Through Computer Vision Methods With Rudimentary Implementation Under Android” [HTML, PDF] about Car Navigation have been uploaded. Due to some server error I am still trying to gather all the code for the former in order to upload it. But that too will come soon.
ed and green hues represent the matching from dual-scale classifiers of the rear of cars. Some more bugs were removed in this latest iteration of the implementation.
he circle at the left shows whether the car is getting closer (white) or going away (red). The size of the circle is indicative of length.
Extensive Static Test of Car Tracker (for Navigation/Collision) and Tracking With Increased Sample Size
ASED on further experiments, at the expense of performance in terms of framerate we can easily improve accuracy to the point of perfect tracking for particular cars. This is done by increasing the sample (window) size on which the tree is trained.
Using 20 positives and only 10 negatives for training.
20 positives and 20 negatives with a sample size of 40×40
ased on cascade training with just 10 negatives and 20 positives (various cars and distances, poses). The real FPS rate is 6+; in the demo it’s decreased by the grabbing/streaming of screenshots for video capturing.
Zoom Changes and Local Binary Patterns Classifier Applied to Red Car
Car Tracking Test (Static With Panning)
6 FPS for tracking (1-minute video)
This video shows tracking of a car based on training with just 10 negatives and 20 positives (without cars like the one in this demo). The real FPS rate is around 6; grabbing and saving a video such as this (in real time) entails a massive performance penalty, so this demo cannot show just how smooth the tracking really is. For a classifier trained on more examples the performance will be comparable. If some code cruft is removed and the rendering gets optimised, 8 FPS seems reachable (this device generally captures raw video at about 10 FPS).
ased on a cascade training with just 20 negatives and positives, performance of some merit can be reached (yes, even with a small number of images). Here are three videos that show it in action, where the goal is to track the car (for navigation purposes as will be shown another day).
Local Binary Patterns With Red Car
Local Binary Patterns With White Car
Local Binary Patterns With Nearly Black Car
Note: My program’s demo should be treated as just the testing of a concept at this stage. I will later on proceed to proper training and thus good performance.