1 Year, 1000km: The Oxford RobotCar Dataset Paper Notes
Article Directory
1. Data set overview
1.1 Main information
This paper is also a contribution to the data set. Although there is no methodological improvement, the author has contributed a new data set to everyone.
- Data set name: The Oxford RobotCar Dataset
- Application area: Autonomous driving
- Collection time: May 2014 to December 2015
- Collection location: Route in central Oxford (frequent crossing)
- Quantity: Nearly 20 million images
- File size: The total uncompressed size is 23.15TB
- Purpose: To be able to study how to solve the long-term positioning and mapping of autonomous vehicles in the real-world dynamic urban environment.
- Additional: Collect data under all weather conditions (heavy rain, night, direct sunlight and snow)
- The complete data set download link: http://robotcar-dataset.robots.ox.ac.uk
1.2 Example of 3D visualization of collected data
Figure 1 shows a sample of 3D maps constructed from different regions in the data set:
二、RobotCar platform
1.RobotCar display
Figure 2 shows an overview of the robot car collecting data and the location of the sensors.
x forward (red), y right (green), z downward (blue)
2.2 Sensor equipment
The location of the sensor on the vehicle is shown in Figure 2, specifically:
- 1 x Point Grey Bumblebee XB3
- 3 x Point Grey Grasshopper2
- 2 x SICK LMS-151 2D LIDAR
- 1 x SICK LD-MRS 3D LIDAR
- 1 x NovAtel SPAN-CPT ALIGN inertial and GPS
navigation system
3. Data collection
3.1 Collection method
- During the entire data collection period, the vehicle was driven manually and did not use autonomous functions.
- The main data collection route is shown in Figure 3, which has been traversed more than 100 times in a year (in some sections of the route, significant structural changes occurred during the data collection period)
- The table in Figure 4 below lists the summary statistics of the data set throughout the year, as well as the number of times of traversal of different condition labels.
3.2 Data format
The author divides the data set into separate paths, and each path corresponds to a one-way traversal. In order to reduce the size of the downloaded file, each traversal is further divided into blocks, where each block corresponds to a segment of about 6 minutes of travel. In a traversal, blocks from different sensors will overlap in time; however, blocks do not correspond between different traversals. Each block is packaged as a tar archive file for download; the folder structure in the archive file is shown in Figure 5 below.
Figure 5 shows the directory layout of a single data set as shown in the figure below. When downloading multiple tar archives from multiple traversals, extracting them all into the same directory will preserve the folder structure. The folders beginning with ldmrs_ are 3D lidar scans; lms_front or lms_rear are 2D lidar scans; the ones beginning with momo_ are taken by the surround view camera; the folders beginning with stereo_ are taken by normal cameras; gps.csv And ins.csv is the GPS and inertial sensor data of SPAN-CPT (for details, see the format introduction of the data type in the next paragraph).
In order to facilitate transplantation, the author converts the sensor data in the internal log format into a standard data format. The format of each data type is as follows:
- Image: All images are stored as lossless compressed PNG files^2, using uncorrected 8-bit original Bayer format, the file structure is (camera)/(sensor)/(timestamp).png
- 2D lidar scan: the 2D lidar return value of each scan is stored in a binary file in the form of a double-precision floating point value, similar to the Velodyne scan format in [6]. The structure of the file is (laser)/(timestamp) .bin, where (laser) is lms_front or lms_rear
- 3D lidar scan: The 3D lidar returned by LD-MRS is stored in the same compressed double-precision floating point binary format as the 2D lidar scan. The structure of the file is ldmrs/(timestamp).bin
- GPS+Inertial: GPS and inertial sensor data from SPAN-CPT are provided as csv files in ASCII format. Two separate files are provided: gps.csv and ins.csv files.
- Visual Odometer (VO): The local error in the GPS+inertial solution may cause discontinuities in the local map constructed using this sensor as an attitude source (due to loss of satellite signals or reacquisition). For some applications, it is preferable to smooth the local pose source that is not necessarily globally accurate, such as the local 3D point cloud structure shown in FIG. 5.
to sum up
In this article, the author dedicated a new large-scale data set to us, mainly dedicated to related research in the field of autonomous driving, as well as the experience and lessons the author summarized in the process of making this data set: First, only record raw data , The second is to use a forward compatible format, and the third is to separately record and process data. These three points of experience and lessons are worthy of reference and study by researchers who will make data sets in the future.