Jinglianwen technology provides 3D point cloud image annotation service

Autonomous Vehicle rely on input high-quality training data to make driving decisions , The higher the quality of the data , The more accurate the decision the vehicle makes , The safer the vehicle . And 2D Compared with image data ,3D The point cloud data contains the distance of the target 、 angle 、 Speed and other information , It does not depend on the external lighting conditions or the radiation characteristics of the target itself , Use at the same time 2D Images and 3D Image data can better perceive the road environment around the car , Provide decision-making basis for perception system .

Data annotation is the key to the effective operation of artificial intelligence algorithms , To realize artificial intelligence, we must first let the computer learn to understand and have the ability to judge things , Want to further improve 3D Accuracy of point cloud annotation , It needs a lot of accurate annotation data to support .

One 、 Overall requirements

The road scene collected by laser radar and millimeter wave radar 3d Point cloud data for information annotation .

Two 、 Labeling requirements

1. Type of marking

2. Dimension box size

• The goal itself must all be 3D Frame surrounding , No leakage and no noise .
• If the boundary of the target object is clear , be 3D The distance between the frame boundary and the real boundary of the target subject cannot be greater than 10cm.
• If the point cloud is missing due to incomplete scanning of the target object , The brain needs to fill the gap and lose the boundary , It can be marked according to the following values ：

• Pedestrians with backpacks 、 Umbrellas and other objects , Pedestrian 3D The box needs to contain these small objects .
• When pedestrians are too close , Pedestrian frames may overlap to some extent , It can be marked normally .
• Pay attention to whether the three views Fit ; Pay attention to whether the straight ahead direction is correct .

3. Mark the location

• All target objects on the ground have 3D The bottom of the callout box must be close to the ground , Not above or below the ground .
• If the target object is too far away, the point cloud in the area where the target is located is sparse 、 There is no ground point or it is difficult to determine 3D The height of the bottom of the frame , Refer to the height of the nearest ground point cloud line and the nearest 3D Box to determine the approximate height .

4. Dimension direction

• The point cloud frame shall be marked according to the driving direction of the vehicle .

5. Mark the range

• 100m Radius range .

6. attribute ID

• The same vehicle in the scene 、 obstacle 、 When pedestrians appear many times, their properties ID Be consistent .

7. Special scenes

• If there are some special ⻋ Vehicle hour , Such as water sweeper 、 Cement tank ⻋ etc. , It is necessary to mark according to the shape of the largest circumscribed cube 3D box .
• If there is one ⻋ Divided into multiple parts , For example, dragging and hanging multi section cards ⻋、 A bus with too many stops ⻋ etc. , Each section needs to be ⻋ Mark respectively 3D box .
• Only mark objects with certain positions and shapes , There is no need to frame the obstacles that cannot be determined by image and point cloud .
• If the target is pedestrians pushing themselves ⻋、 Push shopping ⻋、 Push the baby ⻋ When scene is waiting , Pedestrians and ⻋ The frame shall be marked separately .

3、 ... and 、 Acceptance criteria

• Don't miss the box , Error free box , No more boxes , Residual shadows and miscellaneous points do not need to be marked .
• The direction of marking should be correct .
• The frame is required to closely fit the point cloud , Do not omit punctuation , The outer boundary does not exceed the boundary of the clear point cloud 10cm Location , The marking accuracy is 99% above .
•