Product Parameters

The SLAM solution of Eyesight Technology supports both LiDAR SLAM and visual SLAM. At the same time, Eyeview Technology supports the SLAM solution of laser radar+visual fusion, which means using both laser radar and camera for positioning and navigation in SLAM.

Lidar and camera belong to different sensor categories, and they have their own advantages and disadvantages:

The advantage of LiDAR is that it can provide accurate distance information, and because LiDAR emits light on its own, it can work in completely dark environments. The disadvantage of LiDAR is that the sampling points are relatively sparse (such as some single line LiDAR with only 360 measurement points per scan), and the SLAM method of LiDAR depends on aligning geometric point clouds, which may cause SLAM failure in geometrically degraded scenes (such as long and straight tunnels).

The camera can provide rich environmental information (RGB) and the information is relatively dense (for example, a regular monocular camera can provide image information of 640 * 480 pixels). The disadvantage of cameras is that they are sensitive to environmental lighting conditions, such as not working in dark environments. At the same time, in environments without distinctive features such as cloudy days, snow on the ground, and all white walls, visual based SLAM can also face difficulties.

Product Introduction

Visual SLAM is a set of algorithms that use visual sensors for positioning and navigation. Using visual sensors instead of LiDAR can effectively reduce the cost of SLAM products, while maintaining similar accuracy to LiDAR. However, pure visual SLAM is difficult to effectively locate in a single repetitive scene, and the sensor+IMU solution can effectively solve this problem:
  The fusion of vision and IMU can leverage the higher sampling frequency of IMU to increase the output frequency of system vision;
  The integration of vision and IMU can improve the robustness of vision, such as the erroneous results of visual SLAM due to certain movements or scenes;
  The fusion of vision and IMU can effectively eliminate the integral drift of IMU;
  The fusion of vision and IMU can correct the bias of IMU;
  The fusion of monocular and IMU can effectively solve the problem of unobservable monocular scales.
The visual SLAM system supports three types of cameras: monocular, binocular, and depth cameras.

Product Introduction

Implement a tightly coupled framework of laser vision inertial odometry through smoothing and mapping, for real-time state estimation and mapping.
  In the form of a factor diagram, it consists of two subsystems: a visual inertial system and a LiDAR inertial system;
  When two subsystems detect faults or when sufficient characteristics are detected, they can work independently;
  VIS performs visual feature tracking and can choose to use LiDAR frames to extract feature depth information;
  By optimizing the error of visual reprojection and IMU measurement, visual odometry can be used as the initial value for laser radar scanning matching, and constraints can be introduced in the factor graph;
  After correcting the point cloud using IMU measurements, LIS extracts the edge and plane features of the LiDAR and matches them with the feature map maintained in the sliding window;
  The estimated system state in LIS can be sent to VIS for easy initialization;
  For closed-loop detection, candidate matches are first identified by VIS and further optimized by LIS;
  In the factor diagram, the constraints of visual odometry, LiDAR odometry, IMU pre integration, and closed-loop were jointly optimized. Finally, the entire system outputs pose estimation at the frequency of IMU.

The usual SLAM and localization algorithms are based on the assumption of a static environment, which assumes that all objects in the environment are of the first type. Obviously, this is not the case in real environments. The dynamic environment is the true situation of more application scenarios, so solving SLAM in dynamic environments is an important issue that needs to be addressed in the industry.

• During a single mapping process, high dynamic targets are filtered out in real-time to avoid them appearing as unnecessary obstacles in the SLAM generated map. The application scenarios of this feature include situations such as pedestrians, moving vehicles, animals, etc. during the SLAM mapping process.
• During multiple mapping processes, the map is automatically updated to reflect the current status of low dynamic targets in the map, in response to their appearance, disappearance, displacement, and other situations. The application scenarios of this function include adding an obstacle within the inspection area, removing or changing the position of an object in the original environment, and so on.
• During multiple mapping processes, when the robot moves into an area that was not previously mapped, a new map of the area can be added based on the previous map. The application scenarios of this function include the need to extend the inspection area to areas that have not been previously mapped.

Lidar Obstacle Avoidance

During the driving process, the laser radar is continuously used to detect obstacles and feedback the obstacle information to the path planning module. The path planning module plans a path to avoid obstacles based on known map information and real-time obstacle information, and instructs the robot to avoid obstacles.

Ultrasonic Obstacle Avoidance

During the driving process, ultrasonic waves are continuously used to detect obstacles and feedback the obstacle information to the path planning module. The path planning module plans a path to avoid obstacles based on known map information and real-time obstacle information, and instructs the robot to avoid obstacles.

Machine Vision Obstacle Avoidance

During the driving process, visual perception is continuously used to detect obstacles and calculate their pose, and then the obstacle information is fed back to the path planning module. The path planning module plans a path to avoid obstacles based on known map information and real-time obstacle information, and instructs the robot to avoid obstacles.