To contact me while I live blog about this, use twitter @SinaBahram. Dean of the college computer science at CSUN is talking. One of the fastest growing comp sci programs in the nation. It’s remarkable, the solutions students have come up with in this space. The microphone is handed over to someone else. I’m not doing names, sorry. Maybe I can fill in later.
- taragne recognition: prevent it from going into mud, sand, etc.
- tree roots, bushes
- manmade obstacles like bunches, trash cans, polls
- Also moving obstacles, like bikes, cars, people, etc.
Today, not able to deal with cars, just yet, but we’re working on that. (got a laugh)
Laser range finder, GPS, camera, etc. The data is integrated into a local map and used as part of planning. The recognition of the environment then allows a plan for movement to be thought of and executed by computer. This is standard autonomous vehicles, 101.
We need computer vision, cognition and planning, motion command generation method e.g. EEG, factial signals, voice activation, etc.
Then of course GPS, and we also need local navigation, not just global navigation. So, if you have an obstacle, you have to make real-time detour decisions, etc.
Another microphone change
Changes to a wheelchair
Added power system such as additional battery, regulators, etc. Mechanically, the wheels got bigger. They are fome filled, not air filled, and that optimized ground clearance and speed and comfort. This was also done with casters on the back. Now another microphone change.
Three types of taragne recognition. Using a camera that runs at 30fps. It has a max viewing angle of 66 degrees, and has white balnace, etc. It’s mounted on top and points forward. Shows an image of typical environment. So, identify grass, pavement, and dirt, and then draw those boudnries, virtually that is, and so the computer distinguishes between these taragnes. So, first we do grass detection. They are doing standard color detection e.g. mixing color channels, going to gray scale, etc. Then they go to dirt detection, and then they can overlay these results, of course, since you have grass nad dirt.
Now, shadows cause issues. So that might block a path. Resolution, they implmeented a shadow filter. It just takes shadow filters from source image and filters them out.
One they have grayscale of grass and dirt areas, with no shadows. They thresholde that image into a binary image that indicates preferable Vs. not preferable for path finding purposes. Then they noise filter that. So it’s just cleanup work. And, then they take the image and do some view dredirecting, so they have a top down image view. Finally, the distance of the increments int eh image is calculated. They then sedn this data into a component for processing. Anotehr microphone change.
How can you perceive and make decisions on those perceptions? We have laser range finder, camera, and gps, so we need to use that to make decisions on where to go.
Laser range finder gets them depth, and so they can use it for how far things are. They are showing data from the finder overlayed with an actual image so the sighted folks in the audience can understand depth detection.
They convert, and it essentially looks like a bird’s eye view. They have a polar histogram of this data for various edge detection and other purposes, and they overlay distances on top of this for boundry detection, I beleive, but I have some questions for him about that later on.
Two modes of operation. Hybrid mode takes commands from an EEG headset. Left, right, stop, etc. It is activating in background to make sure no unsafe movement, etc. Autonomous mode is just it gets one command nad then it just goes that way, avoiding obstacles, etc.
GPS can be used to get them course navigation data.
Radial polar histogram is what they developed, basically. So they are trying to determine radial turning distances and directions, but optimizing obviously for path finding, obstacle avoidance, etc. Chosing a turning direction can be hard. You coudl just say, go in a direction, then try, but these guys are trying to chose which curve to use to get there.
The layout of cognition system takes measurements, creates this histogram, has the local map discussed earlier, and grouping data to determine desired block e.g. path finding by any other name. The math in their algorithm determines boudnries e.g. they have a buffer on the wheelchair, virtually speaking, so in other words, even if they are off a bit, you don’t brush up against obstacles. The velocity and acceleration functions are smooth, so that’s basic calculus and just smoothen it out for the user. So, as the range finder sees nothing in front, it keeps going, and it keeps assessing and sampling and making sure it’s on the right path and that there’s a way to get there. Now it approaches a corner, and it wants to turn, so it choses the right turn radius and avoids the two walls in the hallway or whatever, and he’s hsowing a video of this, I beleive, and then there’s a narrower hallway and there’s closed doors, and so the chair handles that as well. Anotehr microphone change.
Originally they used an EEG headset, which she’s wearing as she talks. There are four commands: forward, left, right, or stop. But she found it difficult to go past two commands, so she wanted alternative user commands. She’s going to tell us about three different command types, EEG commands, speech recognition, and other stuff. I’ll cover as I get there.
They are using the emotiv headset. I’m familiar with it, and I’ll post some URLs later on.
She’s saying it takes in thoughts, but you should know it’s just brain waves, not actual thoughts. Emotiv has 14 sensors, or electrodes, that detect brain waves, at a very course level. I believe they are using the idea called motor imagery. She’s not calling it that, but it simply means, make the user think left really hard, and then the signal gets picked up and it interprets it as left.
You can also blink or do other actions and it can recognize that. They are using the congative suite, an expressive suite, etc. So in other words, moto imagery, fatial expressions, etc. So, the user smiles, and you can recongize that and take an action.
The displacement of the gyro in the headset is also used. Remember, there’s a headset, so the wheelchair can interpret those as movement commands. Tehre’s a GUI. If the user is sitting still, the red dot shows as neutral area, but if they tilt their head right, ro even turn it right, it sends out a signal to turn right. They then take these commands and send them as keystroke commands. Not sure why they bother, since they could natively interpret, no? *remember to ask if I ever gatch my breath*
Then they do speech recognition. They are just using a windows computer, and so they are using built-in microsoft speech recognition software. They are also using labview, just some software (URlL later), and they can process those textual commands form the speech.
She shows a video of someone performing some of the actions we just discussed. Microphone change.
They have standard stuff, compus, GPS, accerometer, etc. They use basic shortest path algorithms to do pathfinding. Now, during this algorithm’s running, they recalculate the weights and pick optimized paths based on motion. Now he’s explaining basic path finding and shortest path. It won’t be explainable in this live blog, but tweet me later, and I’ll help explain, or point to rellavent wikipedia articles. It just sounds complicated but is very powerful and easy and straight forward. Now he’s showing a video again of this in action e.g. path finding. He points out some tough patts RE recognition. You know, some computer vision bugs and performance issues that they’ve worked on. Video goes on. Basically they are showing path finding, obstacle avoidance, etc.
Question: can you go in reverse? Well, it can, but no sensors pointing back. In non-automated mode, it definitely can, in that you can give those commands.
Question: you showed it on a marked path, but what about in just a parking lot or open area? He responds that it’s just shortest pathing it’s way to the thing, but if there are obstacles, it’ll avoid them, but of course, this requires an endpoint, because it needs a goal.
Question: are you just using GPS? Yes, just GPS for location, no dead rekkenning.