In trying to add some self control to my InMoov I work on making it find an ARUCO marker and position itself in front of it using its macanum wheel base.

A disadvantage of the mecanum wheel base are the different possible movements and - depending on the surface - quite a lot of slip which makes wheel encoders  a poor solution.

In addition I have been pointed to and read quite a bit about SLAM and its use in improving navigation skills for the robot. One thing I read is that without odometry only poor results can be achieved. So I started to experiment with images taken from a camera pointing straight down at the floor.

I started to google for existing solutions for this and found that opencv has some nice functions that should allow me to retrieve shift distances between 2 sequential images.

After playing around a bit with images and algorithms I was able to achieve the following:

• Find keypoints and descriptions in image A

• Find keypoints and descriptions in image B

• Call a matching function for these points
• Use the matching key points to calculate and visualize the shift distance of the 2 images

(here the 2 images are shown side by side with connections between the matching points)

Mounting a cam on my cart and driving around immediately showed that my cheap webcam was not up to the job and delivered only slurry images while in movement. With these images I could not get sufficient keypoint matches. Looking for a better cam I found the PS3 Eye which promised  frame rates of up to 60 for 640*480 images. Found a used one advertised and payed 20 bucks for it. Found also a windows driver (3 bucks) and instructions how to set frame rate and image size of the cam.

        cap = cv2.VideoCapture(1)

        cap.set(cv2.CAP_PROP_FPS, 80)

        cap.set(cv2.CAP_PROP_FRAME_WIDTH, 320);

        cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 240);

Results improved but I still had a bit of a timing problem running the code (about 300 to 400 ms per image comparison with my laptop (core I7 7500, Intel HD 620, python 32bit). I found then ORB and bruteForce matching and after many tries with thresholds and other ORB-settings finally have a solutions that uses only 100 to 150 ms per comparison on my laptop (about 7 ms!!! on my main PC with a GTX 1070) and showed rather good results.

orb = cv2.ORB_create(nfeatures=250, edgeThreshold = 4, patchSize=20, fastThreshold = 7,  scoreType=cv2.ORB_FAST_SCORE)

This way I can command the cart to move a certain distance and then stop.

On the cart I also have a BNO055 installed so I know the current orientation of the cart. In addition I am currently trying to take depth values from the kinect and try now to combine these with the position and orientation values from the cart to draw a map of my room.

Initially thought I could find existing python projects the help with creation and update of a floor map but results are rather diffuse. Octomap looked promising but I was not able to find instructions or examples how to make use of it yet.