Introduction : the development of navigation, driver assistance, automatic driving and other technical sophistication on the map put forward higher requirements. Conventional level road map for the many deficiencies of intelligent transportation systems, the demand for autopilot applications, we propose a method of using the technology to produce high-precision inertial navigation visual map.
This article will first introduce the mainstream of visual and inertial navigation equipment, visual INS integration framework and key technologies, high German in generating high-precision maps and road signs computing solutions based on the ground to identify the elements of visual way, concluded the technology in high challenges faced by fine map accuracy and future direction.
Visual inertial navigation technology has broad prospects
High-precision map is one of the core technology autopilot, accurate map of the unmanned vehicle positioning, navigation and control, and security. With the continuous development autopilot, more and more enterprises choose the car and map suppliers. High-precision map need to consider the size of the problem and the real-time, high-Germany can provide large-scale service-oriented data models of different brands in the high-precision map industry with leading edge.
Currently, the high German completed the country more than 32 Wan Gongli high-grade road of high-precision map data, with the use of laser radar acquisition, image visual INS fusion of two ways.
Visual image data collected by way of the inertial navigation binding, one can greatly reduce the cost. On the other hand, based on high-precision map of the visual image has certain advantages in the identification, can improve the efficiency of lane-level elements of the job. Therefore, this technology has a broad prospect in high-precision large-scale map production.
High-precision map by the map information with high accuracy composed element vectors, acquiring the element information with high accuracy maps, one map element target is to obtain a visual image by identifying, on the other hand the vehicle acquired by the position and the attitude with high accuracy inertial navigation information, two fused into a corresponding vector map elements.
INS visual tool hardware articles
Visual equipment
Mainstream visual equipment in accordance with different ways of working, the camera can be divided into monocular camera ( Monocular ), twin-lens camera ( Pay Per View Stereo ) and a depth camera ( RGB-D ) three categories.
Monocular camera simple structure, low cost, disadvantage is that the picture is two-dimensional to three-dimensional mapping plane, the lack of depth information, not to calculate the distance between the objects in the scene with us through individual pictures, only motion in order to estimate the depth.
Binocular camera by a monocular camera composed of two, but the distance (baseline) between each other are known. We estimate the spatial position of each pixel is determined by baseline. Binocular camera to measure the depth range associated with the baseline, the greater the baseline distance, the farther can be measured.
So, no car is equipped with a binocular camera usually is a big guy. The disadvantage is that the configuration and calibration are more complex, and its accuracy is limited by the depth range binocular baseline resolution limit, and calculates a parallax very computationally expensive.
Depth camera principle is infrared light structure, similar to a laser sensor, an active emitting light to and receives light returned from an object, the measured distance between the object and the camera. This part is not as binocular camera, as calculated by the software to solve, but by physical measurements, compared to binocular cameras can save a lot of calculations.
The disadvantage is that there may be a depth camera measuring a narrow range, noise, small field of view, vulnerable to sunlight interference, transmission can not be measured materials and many other issues, the outdoor scenes difficult applications.
For high-precision map requires large-scale production requirements, monocular camera because of its low cost, easy to install features it is the mainstream high-precision map visual equipment.
Inertial navigation equipment
Inertial Navigation System (INS) is not dependent on external information, nor to external radiation energy autonomous navigation system. The working environment includes not only the air, the ground, but also underwater.
The basic operating principle is based on Newton's laws of mechanics INS basis, by measuring the acceleration vector in the inertial reference system, it will be integrated over time, and transforms it into the navigation coordinate system, it is possible to obtain a navigation coordinate system information speed, yaw angle and position, etc., are widely used in military, mapping, resource exploration, robotics, autopilot and other fields.
Inertial navigation system with interference, independent and strong, high-frequency data, and good stability. Press the drift rate from small to large it can be divided into level navigation, tactical, industrial, automotive and consumer level. Currently autopilot and high-precision map making more use of tactical field of inertial navigation equipment to meet the demand for high-precision positioning.
In addition, inertial navigation system has developed a variety of ways flexible INS, INS fiber, laser inertial navigation, MEMS INS like. Wherein MEMS ( Micro-Electromechanical Systems, the MEMS ) having a small size, light weight, low power consumption, low price, impact resistance, etc., are widely used, it has been extended to the field of tactical applications low precision.
Be integrated inertial navigation system error used alone, with the practical application of the multi- GPS and Compass represented GNSS ( Free Join the Satellite Navigation the System, the GNSS ) system and other auxiliary systems constituting the composition, to obtain the global position of the carrier.
When the satellite signal is lost, you can get a more accurate real-time position and orientation estimated by INS integration. That do not require real-time mapping application, by the smoothing algorithm can achieve higher positioning accuracy.
In the field of mobile mapping, another role of the INS and Laser external sensors and cameras. By GNSS pose of the resulting coupling, providing high-precision and high-frequency positioning of laser pulses emitted posture image gesture, through an external calibration between the sensors, the projected information corresponding to the global three dimensional coordinate system.
Another configuration is a combination INS inertial visual odometry (visual sensor coupled to the Visual Inertial Odometry , the VIO ). Vision sensor in rich textured scenes SLAM better, but if they take up less moving objects or features scenes photo subject, vision sensors fail.
Inertial Navigation Data Fusion overall positioning accuracy can be improved, and continuity. MEMS inertial navigation unit is widely present in the smart phone, Apple's ARkit and Google launched the ARcore frameworks provide appropriate VIO implementation to support augmented reality applications.
Multi-sensor fusion navigation program has become a trend, inertial navigation system first and GNSS combination, combined with the integrated navigation system composed of an image sensor, a laser radar is currently the research focus and direction of development of autopilot and map making in the field of high-precision.
INS visual framework and key technologies
Inertial fusion of mainstream visual frame is divided into two parts: the front-end and back-end. A front end for extracting the sensor data to build the model state estimation, the front end of the rear end optimize the data provided, the final output of the camera position, posture, and global map, as shown in FIG architecture:
Visual art inertial navigation frame is a front end and a rear end key optimization techniques, described herein is the use of a sliding window mode local visual integration of the INS relative optimization, when initialization fails purely visual considerations into SFM plus INS alignment manner initialized, there will be a global optimization after a relatively optimized, and finally do the absolute optimization of the entire map.
High German high-precision production technology solutions map
High-precision map production mainly from two types of elements of a class of road signs, such as road guide signs, traffic lights and so on; one is to identify the ground, such as lane dividing lines, arrows and other guides. Two categories of map elements are first calculated position, then the gateway linking the elements and road, get attribute information and geometry information elements.
Production of map features to automate manual tasks and extract blend. First, the image and the trajectory of the data collected outside the industry solver, INS obtain visual information needed automation, the integration of technology generates visual map elements according to the INS, the use of artificial conducted on the basis of automation on the map Web model editing, improve the accuracy of map elements, and finally stored in the corresponding database.
Perception result Example:
Generated map example:
Outlook
基于惯导视觉的高精地图生产方案有很多，国内外公司像Moment、宽凳科技，lvl5等都在研究，但是从目前市面上看，由于设备成本限制，基于视觉的高精地图精度极限在10cm。
后续，基于视觉的高精地图发展可能是朝着多源数据融合的方向，即同一道路多次采集，不同设备多次采集获取的数据源融合在一起，提高精度的同时提高地图更新的时效。
高德扎根于地图行业，有丰富的地图数据源，有行业领先的自动化生产技术和成熟的工艺流程，为未来基于多元视觉惯导融合的高精地图生产打下了坚实的基础，这些都会进一步推动自动驾驶的发展。
<center>关注高德技术，找到更多出行技术领域专业内容</center>