Author: tmedirat

This week, my team and I worked on getting more test images to measure the accuracy of the MATLAB implementations of the color detection and circle detection algorithm. Though we have decided to go with color detection since its preliminary results were more promising than those of circle detection.

I also spent a large chunk of my time writing the design review paper due next week. My teammates and I changed our system design block diagram to indicate the changes we have made to our pipeline, the major change being the fact that all of our code will now be implemented in Python. This was suggested to us after our design presentation. This shift in design will help reduce any possible latency that could have occurred when interfacing between MATLAB and Python, thus, will better enable us to reach our goal of a one second response time. Accordingly, we had to change our schedule to show dependencies and revisions regarding individual tasks.

As for my task assignments,  I will no longer be working on K-means clustering since Prof. Savvides suggested using SVM instead. This is because the color of our red dot remains fairly constant. Since Isha will be working on the Python implementation of color detection, it will be logical for her to use that code to implement SVM as well. Thus, during the following 2 weeks I will start working on creating a GUI that would have the functionalities we want for our final presentation. I will need to reinstall all packages again like Python and PyQt since my laptop’s hard drive was replaced. I will also be helping Suann with the sensor circuit since I am responsible for reading the analog signal in order to realize a threshold that will result in a 100% accuracy of true positives reflecting when a tap occurred.

I am currently on schedule since the initial task of carrying out K-means clustering was an additional task and I was eventually going to start working on the GUI soon.

We did not realize that a raspberry pi does not have the exact same capabilities of an Arduino, in that it cannot read analog input directly. We may have to switch to using an Arduino to read the data from the piezo sensor to mitigate any possible difficulties with this part of the pipeline.

After our design presentation, the TAs and Professor recommended that we implement all our code in Matlab so that we can avoid any delays in interfacing between Matlab and Python. We are porting all our Color Detection code to Python next week.

We will not use Matlab anymore for our CV implementations. This change was necessary to prevent any latency caused by transferring data between Python and Matlab. This means that Isha will need to work on the CV algorithms for another week in order to transfer the code. However, this will not delay our overall progress since this task will be replacing our initial task of creating a Matlab to Python pipeline to transfer data between the two.

Another change is that we will be implementing SVM to perform dynamic thresholding instead of K-means clustering because in our case, the color of the red dot remains fairly constant.

Below is an updated schedule:

 

This week I continued to work on the Determinant of Hessian (DoH) blob detection algorithm. After talking to Prof. Savvides, I realized that I could fix the scale of the blur level, i.e, fix the sigma of the gaussian kernel used to blur the image. Thus, the algorithm I have implemented can be considered more of a “circle” detector rather than a true blob detection algorithm. Regardless, it suits our purposes.

As has been mentioned in earlier posts, Isha and I are using the same set of images to test our algorithms. These test images were taken in an environment that closely represents our demo setup. Below are the results:

Currently, my algorithm works for some images (test 12, 14) and not for others (test 11, 13). Following this, one of the design decisions we have made is to color the buttons of our GUI white and keep the background darker. This seems to be a sound decision given the fact that we will only want to detect the user’s finger when he/she taps on a button.

Apart from this, I also helped my team setup the camera and projector in order to obtain the above-mentioned test images. Further, we all worked on the creating the slides for the design review.

Next week I will continue to test my algorithm on different images to see where it breaks and tune parameters accordingly.  I will also start working on K-means clustering that will make use of both color detection and blob detection. We hope that this clustering algorithm will provide us with robust results.

I am currently on schedule.