Build & Test Prep
Team Vision for Build & Test Prep PhaseThe main team goal for this phase was to complete the remaining designs that were still open at the end of the Detailed Design phase of MSD I. Upon completion of these designs, a Critical Design Review was held to communicate the final designs to the customer.
During this phase, customer and engineering requirements were updated, and draft test plans were created.
Test Plan SummaryThe customer and engineering requirements were updated to reflect changes in the overall design. The engineering requirements were also updated to include more measurable metrics.
The Customer and Engineering Requirements document can be found here: Customer & Engineering Requirements
Updated Customer Requirements
Updated Engineering Requirements
- Ensure test plans are in place
- Ensure that all applicable test standards have been cited (e.g., ASTM)
- Ensure all ordered materials have been received
- Ensure team has space and equipment necessary to begin building and testing
Risk and Problem TrackingThe updated risks document for this phase can be found here: Build & Test Prep Risks
- Safe zone asphalt vs concrete (road vs sidewalk)
- Stability of camera when driving
- Running image segmentation on the APM
- Clear case may cause problems with glare, need to blackout case on outside/inside
- Ensuring that all dataset requirement conditions will be met and classified
- IR sensor tests failed, what will be new primary curb detection sensor
- 3rd option for camera able to detect curbs
- Issues with this involve cost for manual labor and reliability
- 16 channel LiDAR
- Curb falls between sensor arrays at certain
distances when straight ahead
- Enough time to make turns to avoid curb, enough time to stop
- There is a variability of detection distances on side
- Need to understand all of the parameters that go
into the heightmap calculations for detecting curbs
- What are the ideal values for detection, do they change with a change in angle
- What is the spread at our first distance that we
can’t detect a curb (1 point on the x,y
- Way to alter this to take raw height of the single point and modify the heightmap
- Height of curbs in test area change, can we detect all of them with ideal parameters
- Curb falls between sensor arrays at certain distances when straight ahead
- Integrate data from camera and 16-channel LiDAR into costmap generation
- Removing global map from the costmap generation
- Modifying local planner to incorporate existing cart capabilities
- Change chosen destination from user input to most
available free space
- Define what most available free space is, is this the best way for our scope or future teams
Infrared Sensor Testing
Testing was performed on the RFD77402 loT 3D ToF Sensor to determine feasibility. The sensor was mounted to a cardboard box at 9", approximately the hight of where it would be on the cart, and measurements were taken in various settings. After testing the sensor outdoors, it was determined that this particular sensor would not be a viable option for the APM. The infrared sensor operates with a non-modulated wavelength of 850nm, which causes errors when used outside due to the spectrum of the sun interfering with the infrared receiver. For this reason, the furthers planned tests were not carried out due to the immediate failure of the sensor outdoors.
Outdoor Testing in ShadeThis test involved placing the IR Sensor test setup in a shaded area outside. For this test, the test setup was placed about half of a meter away from a shaded brick wall. The results for this test were found to be fairly accurate measurements.
Outdoor Testing in Partial LightThe same test was repeated for lighter conditions, with the test setup kept about a meter away from the wall.
Outdoor Testing in SunlightThe same test was repeated for cases in which the sensor is in an area affected by direct sunlight. The direct sunlight causes the sensor to output errors, such as a readings suggesting the sensor is further from objects than it actually is (causing the sensor readings to max out in most cases), or errors where no measurement can be made due to the sensor receiver being saturated.
Safezone DetectionInitial proof of concept has been created on a ROS testbench. Camera frames are successfully obtained from the USB camera and passed as input to our trained ENet model. The ENet output image shows a clear boundary between safe and unsafe driving regions. A new ENet model will be trained on our newly gathered dataset to improve segmentation accuracy in unforeseen camera frames.
The ENet output image is then warped to remove fish-eye distortion using intrinsic camera parameters found through camera calibration. The undistorted image is perspective mapped to the ground plane directly in front of the cart by utilizing the estimated extrinsic camera parameters. Once the camera mounting position is finalized the extrinsic parameters will be measured. The resulting perspective mapped image gives a bird's eye view of the area directly in front of the APM.
The final perspective mapped image is translated to a pcl::PointCloud object and published for use by other ROS nodes. This output format may be subject to change depending on future navigation work.
All steps in the safezone process have been compiled into a ROS package named "safezone."
In the below images the original camera frame can be seen in the "Original Frame" window, the ENet output in the "ENet Output Frame" window, the perspective mapped image in the "Perspective Mapped Frame" window, and the pointcloud output in the rviz window.
Camera MountFor a weatherproof and secure camera mounting solution, a GoPro case was used as a housing for the camera. A camera mount was designed and fabricated in order to secure the camera in place inside of the GoPro case. Together, the GoPro case and internal camera mount create a weatherproof housing for the camera, while keeping the camera from moving around inside of the case.
The camera mount was 3D printed and must remain within 1% tolerance of nominal values in order to fit inside GoPro case.
The camera mount documentation can be found here: Camera Mount Drawing
Software PlanningA software diagram was created to show each sensor with corresponding ROS packages and overall software flow. The pre-processing stages for the camera are also shown, which are required before the camera system can publish meaningful information to ROS.
Image Segmentation and Perspective Mapping
Design Review MaterialsInclude links to:
- Presentation and/or handouts
- Notes from review
- Action Items
Plans for next phase
- As a team, where do you want to be in three weeks at your next review?
- As an individual on the team, what are you doing to help your team achieve these goals? (Use the individual 3-week plan template for this)