Testing our ESP32 Stereoscopic Camera - How far away is the egg?
ฝัง
- เผยแพร่เมื่อ 30 มิ.ย. 2024
- A follow-up to my previous video where I put together a stereoscopic camera with a pair of ESP32 Camera boards and a serial interface.
In this video, we test the camera and try to measure distance of objects on stereoscopic vision principles.
#esp32 #esp32cam #stereoscopic #esp32project
0:00 Introduction, Objectives and Test Setup
1:39 Theory
5:55 Working up to the Calculations
9:10 Calculations, Results and Conclusion - วิทยาศาสตร์และเทคโนโลยี
![CAMPปลิ้น | EP.77[1/2] ก็ไม่ได้ชอบเอาชนะ แต่เรื่องอะไรต้องแพ้!!](http://i.ytimg.com/vi/JraaBYOd7I0/mqdefault.jpg)
Just take a picture of your graduated rule at 60cm for ex, then calibration will be obvious. You will be ablte to compensate for each and every parameters. As you can see the main concern here is that the 2 camera are not parallels, so you have an error due to your hypothesis of common center, and a minor (2nd order) error too for non plan to plane correction. In addition lens are not purely geometrical (stenope is) so the projection is not proportional.
Another useful trick is to place another green tape line on the table, perfectly parallel to the main one. Prolungation of the 2 lines will give you an infinite point that may replace your off center.
Thanks for this advice. I will keep working on this with these considerations in mind.
Cameras sometimes do not use full (or the same) sensor size/area for different resolutions. Your fixed offset of -22 that gives much more accurate results suggest that this might be the case here.
I.e. your assumption of 2.2um pixel size at full resolution does not directly translate to the lower resolution.
I might be a bit off with what EXACTLY doesn't directly match between the resolutions, but I hope you get what I'm trying to say.
I think I understand exactly what you're saying: let's say full camera resolution is 1600x1200, but to return the 320x240 resolution I want, the camera only uses 1200x800 pixels to scale from - then my pixel-scaling would be way off. In fact, I remember trying to check the datasheet for such information but didn't seem to find anything and so I used the full resolution scaling assumption... I will need to revisit. Thanks for that!
Human vision use binocularity only up to 3 meters. Beyond that our brain use another clues to estimate distance, such as occlusion, velocity, relative size to known objects, and many others. Binocularity is great only at short distances, below 3 meters. So, you could move the cameras appart to achieve better binocularity at larger distances.
I forgot to mention that I use two esp32cam in a small rover, that I build starting from the many esp32cam based rovers that you can find in youtube. I separated the camera up to 20 cm, and the 3D effect is amazing. Also, instead of controlling the rover with buttons on a webpage hosted on the esp32cam, I wrote a small app on MIT App Inventor that use the gyros on a cell phone, so I can move the rover tilting my head calling the urls on the esp32cam. I put the cell phone in a Google cardboard, and I place two webviewer components side by side so I can have each esp32cam view on each eye. Only one esp32cam is connected to the two motors. I fixed the ip of the esp32cams so the App always find them.
Let's make it have eyelids similar to how a windshield wiper cleans away dirt, and then for added human eye realism, introduce some tiny though visible bacteria to swim around on the surface sometimes swimming past the camera lens just like with real human eye behavioral patterns.
The surrounding area of the eye can have a round OLED or E ink surface that can replicate a developing bloodshot look, so if the robot is waterproof and swimming in a pool, it will also show signs of interaction with the pool's chlorine with red eyes. This abilty to go from clear to bloodshot eyes could also be easily applied to creating exacting convincing robotic replica actors to make newer cheech n chong movies!
Party on Garth, party on.
focus on the accuracy of setup and measurement is one way. but with the help of AI, given enough samples it could be achieved through machine learning.
we human eyes are not that accurat instrument either, but our brain compensated for errors since we are baby.
我做过类似的工作，感觉在理论上还深化了一些，希望联系交流。I have done similar work, and I feel that it has deepened in theory, and I hope to contact and exchange.