Posted on

Raymond Ngo’s Status Report for 2/26

For the deliverables I promised last week, I am not able to provide them because the schedule timing has been tighter than I anticipated, and as a result I was not able to properly tune the parameters for the blob detection. The most it could do was to detect the shadow in the corner. Ideally, a blob detector with better parameters is the deliverable by the deadline next week.

The main reason was the tight schedule between the presentation and the design report. During the past week (from last Saturday to 2/26), a lot of work was done on the design presentation, mainly having to communicate between the team members on the exact design requirements of the project. Questions involving the discovery of a temperature probe that could withstand 500F or higher popped up, as well as questions on the best type of object detection. Furthermore, as I was the one presenting the design slides, it was my responsibility and work to practice every point in the presentation, and to make sure everyone’s slides and information was aligned. In addition, as it was not I but Joseph who possessed robotics knowledge, I had to ask him and do my own research on what the specific design requirements are.

This week, I also conducted research on other classification systems after the feedback from our presentation. Our two main issues are the lack of enough images to form a coherent dataset, hence our lower classification accuracy metrics, and a selection of using a neural network to classify images ahead of other classification systems. One possible risk mitigation took I found was using a different system to identify objects, perhaps using SIFT. That, however, would require telling the user to leave food in a predetermined position (for example, specifically not having thicker slabs of meat rolled up).

We are on schedule for class assignments, but I am a bit behind in configuring the blob detection algorithm. I am personally not too worried about this development because blob detection algorithm is actually me working ahead of schedule anyways.

Posted on

Raymond Ngo’s Status Report for 2/19

This prior week I was getting myself acquainted with opencv and its libraries. I successfully made a function that captures webcam data both in a continuous stream and when a function is invoked. I successfully applied the cammy edge filter (for thickness detection) on a captured image and increased its threshold. (proof below) This is necessary for the computer vision part of the project because this will be the primary way to detect meat thickness for the cooking time estimation.

I am actually currently on schedule. Figuring out features of opencv and trying out some of the tools is important before starting on the real work of creating tools for the project. Furthermore, finding limitations of some computer vision methods is important before the design review.

Next week’s deliverables: some rudimentary form of blob detection. Uses opencv to capture and process an image. This is necessary because the action that kick starts the cooking process is a user placing meat in front of a camera, and this requires blob detection to see if an object exists or not.

Posted on

Team status report for 2/12

One risk we have not considered to the extent we should have until feedback from the instructors for other proposals was slack and integration time. For one, our integration between the software controller and the computer vision algorithm and the integration between the software controller and the robotic arm happen almost simultaneously, which creates the issue of integrating 3 items at the same time, greatly complicating the process of integration. Developing the Computer Vision and software interface might take less time compared to the robotic arm, so one idea proposed is to reduce the development time of computer vision and software by a few days or a week and use that extra time to integrate. Furthermore, testing the robotic arm on an actual grill should take place before the integration step, in the case of a serious misunderstanding of the heat tolerances of the arm.  We also made a basic concept drawing of our robotic system.  Here is our updated schedule, mostly seen at the end of March and the month of April (I’m sorry for the image quality it cannot be improved for some reason, but we have made some dates for integration longer and software and CV implementation shorter). 

Posted on

Raymond Ngo’s status report for 2/12

This past week I took a further look at the types of computer vision algorithms needed to complete the thickness estimation project. While initially we decided on using a neural network to determine the type of meat to help find the cooking time, we decided this would not be a good idea, owing to the different colors (from marinating) and the similarities of various types of meat. We would also have issues finding a proper data set to train on.

 

Instead, I looked through the different methods of finding thickness, and the best way seemed to be using the Cammy edge detector function in opencv. The challenge facing the upcoming week would be finding a way of making sure the thickness measurement (most likely in pixels) is accurate. The second issue would be making sure the meat measurement is correctly measured at a similar environment each time. This would most likely be done by having the robotic arm lift the meat to the same location each and every time, with the only variable being the position  the robotic arm grabs the meat at. However, this ignores the really thin cuts of meat. In the coming week, I will discuss the possibility of removing that type of meat from our testing metric completely, given how different it will be from every other type of meat we plan on testing. Included is the image of the outlier meat cut.

 

3 things to consider when you cook meat Korean BBQ Style - Korean BBQ OnlineKorean BBQ Online