Difference between revisions of "2021FallTeam7"
|Line 103:||Line 103:|
==Final Product Example==
==Final Product Example==
== Schematics ==
== Schematics ==
Revision as of 22:52, 10 December 2021
- Andrew Liang (ECE)
- Jiansong Wang (MAE)
- Shane Benetz (ECE)
- Kevin Kruse (Extension/BIS)
Project Overview & Proposal
The goal of this project is to enhance the existing framework of the ROS Navigation system with the integration of a Lidar in order to avoid obstacles while navigating and return to the correct route. In a short amount of words, our upgraded ROS navigation package integrated onto the Jetson Nano powered autonomous car allows it to detect objects in its path and correct its steering while still following a lane detection/guidance model based on onboard camera imaging. The cursed "DNF" should rarely ever happen. Our objective was to make a car that would be able to continue on despite the circumstance, and unless something else ran into the car, it would find its way around things and not just follow a straight line blindly.
More in Depth:
While the car navigates the track using openCV lane detection, the lidar runs in the background scanning obstacles in front of the car in a certain range. Out-of-range detections and long distances are filtered out. Once an obstacle is detected in that range, the car will steer to a direction based on the angle and distance feedback from the lidar. The steering magnitude is determined by a proportional controller. Once the car passes the obstacle and is obstacle is out of range of the scan, the car will keep executing lane detection, detecting the yellow stripes on the track.
ROS Laps Here
Project Schedule / Gantt Chart
Calibrating the existing ROS framework
One especially important thing for our project is a properly calibrated framework.
To achieve this, you can either adjust the controls to make it fit. However, an alternative and more precise solution is to apply a specific color filter. To do this, we first started a live transfer of the current image.
Once we had the image, we opened a color tweezer using Microsoft Word and detected the color of the yellow center track.
The program now gives us the color in RGB color space. Using a converter, e.g. https://www.peko-step.com/en/tool/hsvrgb_en.html we can now convert this color code into the HSV color code used by our OpenCV module. Finally we just have to configure this filter correctly and add some tolerances, because especially in twilight the colors vary. Once this process is complete, we get a perfectly calibrated system, as shown in the image below. We have to remark, that in this picture even the red line on the left is not detected at all (which is what we aimed for). Lastly, we have to mention that this scales of the HSV color space very by application: The website uses scales up 100 or 360, while our robot uses a scale up to 180. Thus, we used the scale of 360 and devided the resulting numbers by 2 to get the desired configuration.
Link to source code found here: Team7_Final_Project
The origin of the majority of this code comes from Dominic Nightingale's project ucsd_robo_car_simple_ros. We borrowed and modified the existing code base environment, added a pre-existing ROS package to communicate with the LiDAR onboard, and added our own steering logic corresponding to both the lane guidance model and the LiDAR model.
LiDAR: Detects objects at 5 degree increments in a cone in front of the robot (-65 to +65) degrees range so as not to detect obstacles on the sides of the robot.
This cone has a range of 0.15 to 1.2 meters
Output: 3 Element Float32 Array: [Distance to object in meters, angle of strongest detection in degrees, object detected flag]
- LiDAR currently works like a directional range sensor
- Maybe with better LiDAR or more detailed point cloud could train an actual object detection algorithm based on a currently existing backbone
Given the three obj detection outputs: We can actually create a steering equation for steering around obstacles concurrently with waypoints.
Baseline steering equation (standard P controller):
Θ = -(kp * errx)
New steering equation (two-part P controller for second sensor input): It’s opposite sign because baseline steers toward the point while we want to steer away from the waypoint (obstacle).
Where kp is steering sensitivity from 0 to 1, errx is distance to center waypoint in meters, objdetected is a 0 or 1 dependent on if there was a detection, Θdetection is the angle of detection, Θmax is 65 for our angle range of detection, and objdist is the distance from the ego vehicle to the detected object.
This equation will not differ from the old equation if there is no detection, but if there is, it will add a scaled constant based on which side the object was detected on (if it’s slightly to the left of the vehicle then steer right and vice versa), the inverse normalized angular distance from 0 to determine how hard the turn should be (turn should be proportionally less if the object was far from center), and the inverse normalized distance from the detected object to make the vehicle turn proportionally less the farther away from the object it is. Because all of these constants are normalized, the final output of the added portion will always be less than 1 as this portion is designed to provide a correction to the camera navigation and not overshadow it completely.
- Adjust the algorithm to smoothen the driving and steering harshness
- Design it to look ahead a few waypoints to profile a path of motion rather than relying on camera to always have the next waypoint in frame (localization and odometry w/motor encoders, IMU, stereo camera navigation for depth estimation a huge plus)
- Implement I and D control logic so it is better at staying corrected to the track in the first place
Final Product Example
Our car can pass obstacles on the way to it's destination:
... and it can even pass other cars!
Links to additional resources (presentations/source code/GitHub/videos)
- Jack Silberman (Professor)
- Dominic Nightingale (TA)
- Haoru Xue (TA)