LOAM study notes

LOAM study notes

paper reading

This paper is writed by PHD Ji zhang and Prof Sanjiv Singh.Paper propose a new way to solve 3D lidar SLAM problem with low-drift and real time.

In my view,this paper mainly combine two algorithms.One algorithm names lidar odometry, processing the feature points by 10hz to get estimate velocity of the lidar.(a IMU can be used to get accurate estimation).Another algorithm names lidar mapping,process more feature points by 1hz to match new cloud points to the map and extent the map.

Step 1:Cloud points registration

• Extract feature points from lidar cloud points by formula (1).maximum c means edge points and minimum c means planar points.
• 
• Process special feature points to avoid points whose surrounded points are selected,or points on local planar surfaces that are roughly parallel to the lidar beams(a),or points on boundary of occluded region(b).
• 

Step 2:Finding feature point correspondence

• For each point in ${\tilde \varepsilon _k}$ and ${\tilde H_k}$ ,find the closest neighbor point in ${\bar P_{k-1}}$
• Use KD_tree for fast index
• For point $i \in {\tilde \varepsilon _k}$,find j in ${\bar P_{k-1}}$ and l in the preceding and following two scans.(j and l are from different scans)
• compute the feature point to line distance or to plane distance by formula (2) or (3).
• 
• 

Step 3:Motion estimation

• The lidar motion is modeled with constant angular and linear velocities during a sweep.First using linear interpolate ways merging the points in a sweep to a time stamp.
• project ${\varepsilon _k}$ and ${H_k}$ to the beginning of the sweep by formula(6)==symbol description 13;
• Combine 2,6 and 3,6 ,get the geometric relationship between the edge or planar feature point and the corresponding edge line or the planar patch.In formula (7)(8);
• Solve formula (9) to get $T_k^L\left( t \right)$

Step 4:Lidar odometry algorithm

• 
• Compute a bisquare weight for each feature points.
• 
• Through formula(10) ,try to find the convergence.From $T_k^L\left( t \right)$ we can easily estimate the lidar motion.

Step 5:Lidar mapping

• 

Symbol description:

1. ${P_k}$:Combined point cloud during sweep k
2. ${\bar P_k}$:The projected point cloud.(Correct the influence of lidar motion during time gap.project a sweep of points to a tome stamp.k-1 sweep points project to time stamp k;)
3. ${\varepsilon _k}$:edge points set during sweep k
4. ${H_k}$:planar points set during sweep k
5. ${\tilde \varepsilon _k}$:projected edge points set
6. ${\tilde H_k}$:projected planar points set
7. $T_k^L\left( t \right) = {[\tau _k^L\left( t \right),\theta _k^L\left( t \right)]^T}$:lidar pose transform between [$t_k$,t];(in sweep k)
8. $\tau _k^L\left( t \right) = {[{t_x},{t_y},{t_z}]^T}$:translation in $L_k$
9. $\theta _k^L\left( t \right) = {[{\theta _x},{\theta _y},{\theta _z}]^T}$:rotation in $L_k$
10. $t_{(k,i)}$:point i’s time stamp in $P_k$
11. $T_{(k,i)}^L = {{{t_{(k,i)}} - {t_k}} \over {t - {t_k}}}T_k^L\left( t \right)$:pose transform between [$t_k$,t_{(k,i)}];
12. $X_{(k,i)}^L$:a point in ${\varepsilon _k}$ and ${H_k}$
13. $\tilde X_{(k,i)}^L = R_{(k,i)}^LX_{(k,i)}^L + \tau _{(k,i)}^L$:corresponding point in ${\tilde \varepsilon _k}$ and ${\tilde H_k}$
14. $R_{(k,i)}^L$,$\tau _{(k,i)}^L$:rotation matrix and translation vector corresponding to $T_{(k,i)}^L$