文件列表:

RoboticsND-Project1 (0, 2017-10-11)
RoboticsND-Project1\.DS_Store (12292, 2017-10-11)
RoboticsND-Project1\code (0, 2017-10-11)
RoboticsND-Project1\code\.DS_Store (6148, 2017-10-11)
RoboticsND-Project1\code\.ipynb_checkpoints (0, 2017-10-11)
RoboticsND-Project1\code\.ipynb_checkpoints\Rover_Project_Test_Notebook-checkpoint.ipynb (2188848, 2017-10-11)
RoboticsND-Project1\code\.ipynb_checkpoints\Untitled-checkpoint.ipynb (72, 2017-10-11)
RoboticsND-Project1\code\Rover_Project_Test_Notebook.ipynb (2181929, 2017-10-11)
RoboticsND-Project1\code\decision.py (4011, 2017-10-11)
RoboticsND-Project1\code\drive_rover.py (8183, 2017-10-11)
RoboticsND-Project1\code\perception.py (13827, 2017-10-11)
RoboticsND-Project1\code\supporting_functions.py (7342, 2017-10-11)
RoboticsND-Project1\misc (0, 2017-10-11)
RoboticsND-Project1\misc\Rock_color_analysis.jpg (75376, 2017-10-11)
RoboticsND-Project1\misc\color_map.jpg (282967, 2017-10-11)
RoboticsND-Project1\misc\example_grid1.jpg (7370, 2017-10-11)
RoboticsND-Project1\misc\rockPixels.jpg (64823, 2017-10-11)
RoboticsND-Project1\misc\rock_thresh.jpg (280910, 2017-10-11)
RoboticsND-Project1\misc\rover_image.jpg (16352, 2017-10-11)
RoboticsND-Project1\misc\video_screenshot.jpg (1469327, 2017-10-11)
RoboticsND-Project1\output (0, 2017-10-11)
RoboticsND-Project1\output\.DS_Store (6148, 2017-10-11)
RoboticsND-Project1\output\test_mapping.mp4 (1331126, 2017-10-11)

[](https://www.udacity.com/robotics) ## Project 1: Search and Sample Return ###### Udacity Nanodegree ###### June 2017 [//]: # (Image References) [image1]: ./RoboticsND-Project1/misc/rockPixels.jpg [image2]: ./RoboticsND-Project1/misc/Rock_color_analysis.jpg [image3]: ./RoboticsND-Project1/misc/rock_thresh.jpg [image4]: ./RoboticsND-Project1/misc/color_map.jpg [image5]: ./RoboticsND-Project1/misc/example_grid1.jpg [image6]: ./RoboticsND-Project1/misc/video_screenshot.jpg ### ### ### ### Overview ###### The goal of this project is to program a rover to autonomously navigate terrain, find rock samples, and to retrieve them. The project is based on Nasa's Sample Return Contest. ### *We operated the rover through RoverSim and carried out the majority of the operations in Python.* ### This project will cover the three primary areas of Robotics: *Perception, Decision Making*, and *Taking Action*. ### **Perception**: Our rover is given sight by the front mount camera. The majority of this analysis takes place in `perception.py` **Decision Making**: We decide which actions to take by analyzing our camera and sensor data given by the rover, and testing various conditional statements. This analysis is located in `decision.py` **Action**: We execute our actions by communicating with the rover using: `drive_rover.py` ### Notebook Analysis Our first task of this project is to take in camera images, and transform them into a useable format. This will be completed in four steps: 1. Warping: converting the front mount image to a bird's eye view perspective 2. Thresholding: separating pixels of interest from others 3. Mapping to Rover Coordinates: maps the warped and thresholded image in rover coordinates 4. Mapping to World Coordinates: rotates and translates the image to account for yaw, and its location on the map **Warping** is done by leaning on the computer vision powerhouse that is `opencv`. All we have to do is map where our ground pixels are located so that when the warp is executed, they scale correctly. This is done by placing a grid on the ground in front of the rover camera. ![alt text][image5] **Thresholding** is carried out by `color_thresh()` which shows the navigable terrain. The next task is to understand how to identify rocks. Using one of the rock sample images, I performed a pixel analysis to determine the appropriate thresholds for Red, Green, and Blue pixels that create the yellow of the rock. ![alt text][image2] Then extracted the rock Pixel Images: ![alt text][image1] Then I ran basic statistics on the pixels determining the range and mean of the RGB values of yellow rocks. This turned out to be: `rgb_thresh=(130,105,50)` I then adopted the thresholding logic from `color_thresh()` and applied it to the function `rock_thresh()`, yielding: ``` above_thresh = (img[:,:,0] > rgb_thresh[0]) \ & (img[:,:,1] > rgb_thresh[1]) \ & (img[:,:,2] < rgb_thresh[2]) ``` Which generated the following: ![alt text][image3] This same approach was then applied to obstacles by developing the function `obstacle_thresh()`. Putting this all together allows for us to create a warped and thresholded image to display in the lower right hand corner of the video, I created the function `color_map()`. This function takes in an image, and segments the pixels into the categories: navigable area, obstacles, and rocks. ![alt text][image4] I then placed this image in the lower right hand corner of the video. **Mapping to rover coordinates** is done with the aptly named `rover_coords()` function. ``` def rover_coords(binary_img): binary_img = binary_img[len(binary_img)//2 : , :] ypos, xpos = binary_img.nonzero() x_pixel = -(ypos - binary_img.shape[0]).astype(np.float) y_pixel = -(xpos - binary_img.shape[1]/2 ).astype(np.float) return x_pixel, y_pixel ``` The first line of the function is nonstandard, as I had to alter it to improve `fidelity` during autonomous navigation. It essentially truncates the area in front of the rover by half. I talk a bit more about this below. **Mapping to world coordinates** is done by first a rotation, followed a translation. The `pix_to_world()` function combines both the rotation and translation The video that is generated and included in the project repo in `./output/test_mapping.mp4` harnesses all four of these steps: ![alt text][image6] ### Autonomous Navigation and Mapping `decision.py` was only slightly modified to have the rover brake when a rock was nearby. The results of this are mixed. The rover must be very close to the rock samples in order to trigger the `near sample` flag. I also attempted a conditional steering statement, but again, I did not find much success with it. #### Launching in Autonomous Mode RoverSim Settings `1280x800` `Quality: Good` `35 FPS` ### Approach, Techniquies, and How to Improve Initially, I struggled much with the `fidelity` score--the rover wouldn't make it above 20%. These inaccuracies are mostly caused by roll and pitch angles not being equal, or near, 0. To remedy this, I shortened the amount the rover was looking out forward. I did this alteration in function `rover_coords()` by shortening the number of rows to half of the total length of the image, like this: ``` binary_img = binary_img[len(binary_img)//2 : , :] ``` This alteration immediately improved the `fidelity` score, and was consistently above 60%, `mapping` reached >80%, and usually 3-5 rocks were found and one was usually picked up. *There are many things needed to improve this code. Some of the noteable issues are:* When `time==1000s` the rover eventually gets stucks beneath one of the rocks in the middle. This is usually remedied by turning to +/-15 degrees for a few seconds to induce a 4-wheel turn. The rover doesn't currently drive towards rocks. This can be remedied by using a conditional statement and determining where the rocks are with `rock_angles` and `rock dist`

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：