Don't watch the clock, do what it does. Keep Going. Sam Levenson
Idea Transcript
Greg – Spatial Information Science and Engineering
Deep Learning- self driving car (hobby) Posted on April 10, 2017
Really it is a Tesla.. Four electric motors, batteries, sensors, and two cameras.
Okay…kinda like a Tesla. I have been fasinated with neural networks going back to the early 90’s when I was doing work on forms recognition and hand writing analysis. The idea lost appeal for a long time but has had a resurgence as “deep learning” is being used for processing large amounts of data. Self driving cars is one area where they are making gains. Recently I watched the series by Dr. Lex Fridman(MIT 6.S094: Introduction to Deep Learning and Self-Driving Cars). Besides covering a lot about neural networks he talked about how Tesla instruments their cars to “learn”. Is the car really learning to drive? Not exactly. By driving the car around it gathers data that can be used later on. The data includes images, sound,temperature, GPS and driver reactions. All of this data is feed into a neural network such as tensorflow. The car knows only what it has “seen” before. The system is memorizing every possible situation. Anything that occurs out of context from what it knows could cause trouble. But the more data that is gathered the less likely it will be that something unforeseen will occur. Of course AI is always improving and at some point will be able to make better choices – zero shot learning. I wanted a way to experiment with this myself. Buying a Tesla is out of the question. I could add sensors to my car but that is just asking for distracted driving. Also, I work remote and don’t drive alot. The next best thing would be create a small ‘car’ that I could use to gather data. This car is not going to be on a road. Which means it wont have things like lane lines to guide it. I might build a track where it could be driven. Or just wander around the house and scare the dog and cat.. The picture above shows a small RC car. It has a motor for each wheel. Steering is similar to tank driving. Turns are done by slowing down one set of wheels while speeding up another. Sharper turns can be done by reversing the wheel instead of just slowing them down. Its not very smooth but it gets the job done. The car will is first configured for data gathering. I am using an Arduino with blue tooth for communications. I wrote a simple app for my Android tablet. There are six ultra sonic sensors for determining distance. I have two cameras(only one in the picture) mounted on the front. These will record stereo images which will help determine depth. The first thing I learned is that the ultra sonic sensors will only see a small portion of what is in front of them. The first trial run they completely missed the table or chair legs. The sensors need to sweep the area in front so as to create a point cloud. For this I am adding a pan and tilt control to the sensor mount. Two servos will move the sensor array. Data is being recorded to a flash drive. Data I am recording the data at one second intervals. The car doesn’t move very fast so this rate should be sufficient. The value at each sensor, two camera images and the drive command are recorded. Currently I am in practice mode, refining the app to better control the car. I found some sensors for the wheels to detect the speed of rotation. I think I’d need to upgrade the Arduino to add any more devices so I’ll leave them off for the time being. More later…
Related
Internet of Related Things IORT
back to the jobs data..
About gricker Living and learning View all posts by gricker Õ
This entry was posted in deep learning. Bookmark the permalink.
Greg – Spatial Information Science and Engineering Blog at WordPress.com.