{"id":356,"date":"2020-04-03T04:04:33","date_gmt":"2020-04-03T04:04:33","guid":{"rendered":"http:\/\/mrsdprojects.ri.cmu.edu\/2020teami\/?page_id=356"},"modified":"2020-12-18T18:47:16","modified_gmt":"2020-12-18T18:47:16","slug":"detection-subsystem","status":"publish","type":"page","link":"https:\/\/mrsdprojects.ri.cmu.edu\/2020teami\/detection-subsystem\/","title":{"rendered":"Detection Subsystem"},"content":{"rendered":"

Fall Semester Implementation Details (Final Version)<\/h3>\n

Any autonomous car is required to adhere to the rules of the road, namely staying safely within the desired lane. We wanted our system to follow these rules as well. In order for our lane following functionality to work, we had to make one assumption about the garage environment. We assumed
\nthere to be a solid line dividing the two lanes within the garage, which we saw in our testing environment in the CMU East Parking Garage. We performed lane detection to segment the lane in camera images obtained from the d435. This information is later used in the goal generation and
\nnavigation stage.<\/p>\n

At this point, we have decided to switch from deep learning neuron networks to classical computer vision methods for both hardware and simulation. It turns out that it worked pretty fine in both cases.<\/p>\n

We utilized classical computer vision methods to detect the center line in the lanes. We had initially been pursuing deep-learning-based methods for lane detection in the spring semester. However, we abandoned that method due to insufficient labeled data for training the system to work in the physical parking lot, as well as the fact that the training we performed on simulation data did not translate well to the physical environment. To make sure our subsystem is robust, we manually place a yellow seat belt on top of the center line in the CMU parking lot, since the existing road paint was fading. Our lane detection algorithm is described in five steps, summarized below.<\/p>\n

1. Perform HLS thresholding on a region of interest in the bottom-left of the image in order to
\nfilter out the yellow seat belt
\n2. Detect the line using a Canny edge detector and probabilistic Hough transforms.
\n3. Project the pixels on the line into a 3D point cloud using the depth camera
\n4. Find the normal vector of this new linear point cloud using singular value decomposition (SVD)
\n5. Generate the right boundary line by extending mirroring the current point cloud along with a pre-determined lane width.<\/p>\n

The point cloud of the center line and the right lane boundary is sent to the navigation subsystem as an obstacle constraint in the costmap. We also generated an ego-line point cloud by extending the center line point cloud to the right with 60% of the lane width. This ego-line point cloud is later
\nsent to the goal generation subsystem as an input, and goals are generated along this line. Overall, the lane detection subsystem is crucial to the success of the whole pipeline since both the goal generation and navigation subsystems rely heavily on accurate lane detections.<\/p>\n

Spring Semester Implementation Details<\/h3>\n

The purpose of the detection subsystem is to identify lanes in the environment in order to impose a constraint on where the robot can move. As of now, this system is still in progress, as we have been trying several different methods for lane detection with varying results.<\/p>\n

Hardware<\/h4>\n

Progress up to Progress Review 9<\/h5>\n

We are using classical computer vision methods here. The reason we gave up leveraging deep learning techniques to detect lanes in real-world data was because of insufficient labeled data for training. Moreover, the model trained in the simulation could not be transitioned to the real world properly.<\/p>\n

To make sure our subsystem is robust enough, we manually place a yellow seat belt on top of the faded center line in the CMU parking lot. Our lane detection pipeline in hardware could be decomposed into the followings:<\/p>\n

    \n
  1. HLS thresholding with the image in order to filter out the yellow seat belt.<\/li>\n
  2. Detect the line with opencv<\/code> functions such as Canny()<\/code> and HoughLinesP()<\/code>, as shown in Fig. 1 & 2.<\/li>\n
  3. Project the pixels on the line back to the 3D world with respect to the depth camera.<\/li>\n
  4. Find the normal vector of the line using SVD.<\/li>\n
  5. Interpolate the right line by extending each point cloud with a fixed lane width, as shown in FIg. 3.<\/li>\n<\/ol>\n