Team 31 CART Update (4/3/2020)

Thanks for reading our blog! If this is your first time reading our blog, we recommend you start with our first blog post.

Progress since our last post (March 20 - April 3)

One of the features we’d like to implement on the golf cart is an “environmental-awareness” factor. We have a depth map of the environment that we process to create our driving commands, but we also have a RGB video feed that we can sample. Thus, we plan to feed that into an image classifier to detect people or other mobile obstacles and scale down our max speed. For example, if surrounded by a crowd of people, we’d like to reduce the max speed from 10 mph to 5 mph. We did some tests with yolov3 ("you only look once"), and results were promising, albeit slow on a CPU. Every person in every image was detected, and no objects were classified incorrectly (although our tests weren’t difficult). Tests averaged ~21 seconds to classify an image at 3.4 GHz, on the pre-trained network without adjusting anything. According to the website, the GPU version is up to 500x faster, but we have not yet tested running this on the Jetson Nano, which should be able to take advantage of some of that performance increase.

Output from yolov3, correctly identified a bench

Map and occupancy grid creation using data from RealSense is being done using Real-Time Appearance Based Mapping (RTAB). Using the rtab-ros module, the RealSense is able to create a detailed map in 5-10 minutes (one-time), which can then be used in real time. Further development of the map will result in a 2D (or 3D, which we probably won’t use) occupancy grid that can be used for obstacle detection. Other possibilities with the RTAB module include appearance based classification and localization.

The main challenge for this work period has been the burnt out motor driver. The dual H-bridge motor driver that was controlling the steering and accelerator actuator motors overheated and burnt out, halting testing until it's replaced. A replacement is in the mail set to arrive April 5th. In the meantime, a "bandaid" fix is in place in which we installed a lesser amperage rated motor driver to facilitate small scale testing in the interim. This is only a minor setback that has not delayed the testing schedule at all.

Burned out H-bridge

Another challenge we've faced is the campus closures due to COVID-19. This has eliminated access to the workspace we've been in for the year and has caused us to relocate the CART to a team member's house to continue testing and validating for the remainder of the project. Additionally, as mentioned last time, we have been unable to acquire the 1D laser range finders.

CART loaded up and ready to leave UH


Planned work for the near future (April 3 - April 17)

We plan to flesh out the 2D (or 3D) occupancy grid using RTAB. We also plan to perform obstacle detection tests at home, which may be more difficult than testing on campus, but we should be able to get this done. We'd also like to explore appearance based classification and localization using rtab-ros.

Additionally, we need to finish getting communications set up between the Arduino and the Pixhawk, as well as the Pixhawk and the Jetson Nano. The former is nearly done, but some more code needs to be written and tested. The latter has been on hold, since we currently do not have the Jetson Nano, but we should be receiving it within the next week or so.

Comments

Post a Comment

Popular posts from this blog

Team 31 CART Update (3/20/2020)

Image
Thanks for reading our blog! If this is your first time seeing on our blog, we recommend you start at our first blog post . Progress since our last post (March 6 - March 20) According to our last blog post, we planned to mount all of our sensors onto the CART and have them operational. However, there have been numerous setbacks. The first (planned) setback was spring break. After a grueling week of midterms and project, we took a break and spent some time recuperating. As a result, productivity was low from 3/9 to 3/13. In addition to taking a break, we had other responsibilities ranging from job interviews to thesis projects. However, we planned to increase the amount of time dedicated to the CART (starting March 16th). Unfortunately, as you are no doubt aware, the COVID-19 outbreak was declared a pandemic, and the situation evolved rapidly. Over spring break, the university went from one extreme to the other - at first, we thought no classes would get cancelled. Now, all of o...
Read more

Team 31 CART Update (01/31/2020)

Image
Thanks for reading our blog! If this is your first time seeing on our blog, you might want to read our first blog post . Progress since our last post It has been a while since we last posted. Since our last post, we’ve ordered and received many of our components. We had planned to work on the CART over the break, but due to delays in the procurement process, many of our parts didn’t arrive until the end of the break. Unfortunately, this meant that we were unable to get much work done during the winter break. We’ve hit the ground running this semester. So far, we’ve installed an upgrade to the braking system to allow an additional (more rapid) braking mode to be activated in the event of an emergency. This is a necessary safety feature before we can begin any autonomous driving tests as the pre-existing braking was insufficient for emergency situations. We have also replaced the pre-existing remote controller with a standard COTS 3-channel transmitter/receiver and tuned the motor ...
Read more

Team 31 CART Update (02/14/2020)

Image
Thanks for reading our blog! If this is your first time seeing on our blog, you might want to start by reading our first blog post . Progress since our last post Since the last post, we’ve fully pivoted away from using 3D LIDAR. Feel free to read our last blog post for more information, but the key point is that we don’t have enough money to buy a 3D LIDAR (but we hope future teams will be able to purchase and use the Velabit when it is released later this year). Instead, we’ll utilize depth cameras and high precision, 1-D laser scanners. We have a free depth camera (an Intel RealSense , courtesy of IEEE Makerspace) and a couple of loaner laser scanners (courtesy of the NCALM ). We were able to test the RealSense earlier this week. After reconfiguring the camera for a couple of hours, we were able to get indoor depth readings for distances of up to 23 feet, (~7m) which is great, but below our goal of 10m. On a bright but cloudy day, we were able to get depth readings of up to 40 f...
Read more