Data Sources - AmyPhung/ml_person_detection GitHub Wiki
Previous: Home
Training Data: Waymo
This project uses extensive training data derived from the Waymo Open Dataset (waymo-od). To quote the dataset's "About" Page:
"The Waymo Open Dataset is comprised of high resolution sensor data collected by Waymo self-driving cars in a wide variety of conditions. We are releasing this dataset publicly to aid the research community in making advancements in machine perception and self-driving technology."
Waymo-od contains massive numbers of 20 second data collection runs containing both lidar and camera data. For the purposes of applying this project to the Olin Robotics Lab's autonomous tractor, the project focuses solely on the lidar data, which is nonetheless extensive. The lidar scans are generated from a suite of 4 short-range lidar and one mid-range lidar, and create a detailed 360 degree pointcloud map of the surroundings of the car for up to 80m, in some instances. An example data frame is shown in Image 1. More detailed statistics can be found via the link above.
| 191212_ml_waymo-ex.png |
|---|
| Image 1: Example Data Frame |
In addition to the raw data, Waymo-od also provides hand-labeled bounding boxes identifying four classes of objects in each scene: Vehicles, Pedestrians, Cyclists, and Signs. These bounding boxes fully define the boundary of clusters of points making up each object, and are generated for over 1000 data segments. Each segment of the dataset is taken in a different location, and has different densities of each class, as displayed in the following graphs. Image 2a shows the number of each class type in the training_0000 dataset, where each object in each frame is counted as unique. Image 2b shows the overall count of unique objects in the dataset based each bounding box's object ID.
[191212_ml_graph-class-count.png]] ](/AmyPhung/ml_person_detection/wiki/[[191212_ml_graph-unique-id-count.png)
:------------------------------------:|:-------------------------:
Image 2a: training_0000 Class Count | Image 2b: training_0000 Unique ID Count
Test Data: Panasonic Lidar
The next steps for this project are to apply the classification algorithms developed on Waymo-od to real data gathered from the Olin Robotics Lab's autonomous tractor. The platform has been entered into the Panasonic 3D-Lidar Challenge, and is equipped with a 3D lidar positioned to detect obstacles in front of the vehicle while simultaneously mapping the driving surface in front of the vehicle. Some example data from the system is shown below.
Note that this example is not reflective of the full capabilities of the sensor system provided by Panasonic.
