In this Project the aim was to lower our RMSE (Root Mean Squared Error) to get a better estimation of the location of a bicycle using CTRV motion model via the reading we get from both the Laser and Radar sensors.
Here I tuned the acceleration and the angular noise variables and monitored the RMSE changes. I started with 0.231 as the initial value for the acceleration noise as it is the average acceleration of a bicycle in urban setting, and for the angle 0.7 which is 45 degrees as a starting point. At first I used 9 which is much better than the 30 that was there before, but its still too high for a bicycle. I tried many variations before I ended up with acceleration 0.331 and angle 0.5 which gave the best RMSE.
Upon further examining the RMSE I found out that increasing the acceleration noise while keeping the yaw noise constant increased the RMSE, so every time I increased the acceleration noise I had to decrease the yaw noise which makes sense for the a moving object turns more quickly if it increases its acceleration. I started with the values of 3 m/s^2 for the acceleration noise and \pi / 4 and kept adjusting both values till I was able to lower the RMSE significantly and the final values were 0.8 m/s^2 for the acceleration noise and \pi / 4.
Here I evaluated my filter using NIS for both sensors to check its robustness with different process noises as you can see below:
| value | Laser | Radar |
|---|---|---|
| acc = 0.8, angle = pi/4 |  |
 |
| acc = 0.3, angle = 0.5 |  |
 |
| acc, angle = 9 |  |
 |
| acc = 9, angle = 0.7 |  |
 |
| acc = 0.2331, angle = 0.6 |  |
 |
As we can see both the Laser and the Radar NIS are within the confidence zone, our distribution does not over or under estimate and as a result we get a healthy distribution save for a few values. The best distribution was for the values 0.8 m/s^2 for the acceleration noise and \pi / 4.
| Laser | Radar |
|---|---|
|
|
All of the RMSE values increased save for vy.
All of the RSME values increased, but the vx value had a big increase of 0.4 vs both sensors.
The sensor fusion RSME gives the best result just as the case of the Extended Kalman Filter.
After comparing both RMSE I found that the UKF far surpassed its smelly counterpart, and there is no doubt that the unsencted filter is the better filter.