semantics SLAM Framework to take advantage of hybrid edge The information is on the mosaic .
To extract from the mosaic edge and free-space outline , Different segmentation methods are designed to remove noisy images glare edge Edge of strong light and Twisted edges of objects ( Yes IPM Caused by the ).
Because only freespace Segmentation needs training , Our approach has been reduced labeling burden Label burden .
At the same time, it builds Semantic edge point cloud image and occupancy grid map.
A combination of Unsupervised edge monitor and A group of IPM Edge segmentation method based on . Our approach only needs to be crude freespace mark .
2. Related Work
A. Visual SLAM for Multi-Camera Systems
B. Semantic Visual SLAM in AD
We design a hybrid edge extraction method , Just rough freespace mark , 10X fast .
C. AVP Applications
- Bird's-eye edge extraction: Mosaic and generated freespace The pictures are all input ; And then test raw hybrid edges, I"M Segmentation to remove most of the edge noise
- Mapping: Generate global edge point cloud image , and occupancy grid map
- Odometry: The current... In the local map pose It's done with semantic point cloud registration , The current transformation of the given wheel speed is the initial value
4. Bird's-eye Edge Extraction
There are a lot of mosaics road markings, It's all good information , But the image is polluted by strong light ; \
There are two sub modules , 1. raw edges testing ; 2. The result of removing noise and distortion
A. Raw Edge Detection
This can be done with traditional edge detection ( such as Canny edge); stay freespace With the help of segmentation , Remove the edges off the ground ;
B. IPM-based Edge Segmentation
One of the basic four ways to clear the edges is utilize IPM Inside distortion effect Characteristics of ; Most of the edges are stable radial Of , It's going through every camera focal point Of .
5. Semantic Odometry and Mapping
A. Local Map Generation
The segmentation module can't completely remove the noisy edges . meanwhile , some road markings Or the edge of the parking spot may be removed by mistake , If they happen to be in The direction of the ray On . So the current extraction edge It could be incomplete, unstable .
We built local edge map, Pictured above , The fusion result of different frames is probability .
B. Pose Estimation
First, project the current frame observation to local map ( adopt T), Data association through nearest neighbor search , Current pose :
C. Global Mapping