文件列表:

.DS_Store (10244, 2017-12-01)
CMakeLists.txt (604, 2017-12-01)
cmakepatch.txt (1102, 2017-12-01)
images (0, 2017-12-01)
images\dataset1.jpeg (83701, 2017-12-01)
images\dataset2.jpeg (80414, 2017-12-01)
images\laser_nis.jpeg (76830, 2017-12-01)
images\radar_nis.jpeg (59353, 2017-12-01)
install-mac.sh (512, 2017-12-01)
install-ubuntu.sh (326, 2017-12-01)
src (0, 2017-12-01)
src\.DS_Store (6148, 2017-12-01)
src\Eigen (0, 2017-12-01)
src\Eigen\Array (304, 2017-12-01)
src\Eigen\CMakeLists.txt (607, 2017-12-01)
src\Eigen\Cholesky (775, 2017-12-01)
src\Eigen\CholmodSupport (1670, 2017-12-01)
src\Eigen\Core (12826, 2017-12-01)
src\Eigen\Dense (122, 2017-12-01)
src\Eigen\Eigen (37, 2017-12-01)
src\Eigen\Eigen2Support (3295, 2017-12-01)
src\Eigen\Eigenvalues (1394, 2017-12-01)
src\Eigen\Geometry (1605, 2017-12-01)
src\Eigen\Householder (580, 2017-12-01)
src\Eigen\IterativeLinearSolvers (1594, 2017-12-01)
src\Eigen\Jacobi (645, 2017-12-01)
src\Eigen\LU (983, 2017-12-01)
src\Eigen\LeastSquares (712, 2017-12-01)
src\Eigen\MetisSupport (697, 2017-12-01)
src\Eigen\OrderingMethods (2189, 2017-12-01)
src\Eigen\PaStiXSupport (1467, 2017-12-01)
src\Eigen\PardisoSupport (864, 2017-12-01)
src\Eigen\QR (926, 2017-12-01)
src\Eigen\QtAlignedMalloc (637, 2017-12-01)
src\Eigen\SPQRSupport (930, 2017-12-01)
src\Eigen\SVD (858, 2017-12-01)
src\Eigen\Sparse (594, 2017-12-01)
src\Eigen\SparseCholesky (1433, 2017-12-01)
... ...

# SDC P7: Unscented Kalman Filter [//]: # (Image References) [image1]: ./images/dataset1.jpeg "Dataset 1" [image2]: ./images/dataset2.jpeg "Dataset 2" [image3]: ./images/laser_nis.jpeg "Laser NIS" [image4]: ./images/radar_nis.jpeg "Radar NIS" ## Description: This program implements an Unscented Kalman Filter, with sensor fusion, that uses lidar and radar data to track a moving object. CTRV is used to model object motion with the following state parameters: px, py, v, yaw, and yaw rate. The program works in conjuction with [Udacity's SDC term 2 simulator](https://github.com/udacity/self-driving-car-sim/releases). The simulator provides the program with simulated lidar and radar data. Using this data the program estimates the state of the object being tracked. The estimated state is compared with ground truth values to calculate the root mean squared error (RMSE). --- ## Dependencies: * uWebSocket * CMake * Make The install-mac.sh and isntall-ubuntu.sh files can be used to install project dependencies. For further details see [Udacity's project repository](https://github.com/udacity/CarND-Extended-Kalman-Filter-Project). --- ## Build Instructions: 1. Download Udacity's SDC term 2 simulator. 2. Ensure the dependencies are isntalled. 3. Clone this repository and navigate to its local directory. 4. Create a build directory. ```shell $ mkdir build ``` 5. Initiate CMake from build directory. ```shell $ cmake .. ``` 6. Make the project. ```shell $ make ``` 7. Run ExtendedKF ```shell $ ./UnscentedKF ``` 8. Launch Udacity's SDC simulator and start project 2 simulation. --- ## Results ### RMSE Values The table below shows RMSE values on Dataset 1 in the simulator. The first three columns show values obtained using the Unscented Kalman Filter for their respective sensors (combined uses both lidar and radar). The last column contains RMSE values from the Extended Kalman Filter build in the previous project. | | Laser Only | Radar Only | Combined | EKF | |:---:|:----------:|:----------:|:----------:|:---------:| | px | 0.1026 | 0.1509 | 0.0670 | 0.0950 | | py | 0.0***4 | 0.2038 | 0.0836 | 0.0857 | | vx | 0.5121 | 0.3079 | 0.2846 | 0.4167 | | vy | 0.2370 | 0.2460 | 0.1915 | 0.4228 | ### Simulation The image below is a screenshot from the simulator using the Unscented Kalman Filter from this project. ![alt text][image1] ## Discussion From the table above we can see that the Unscented Kalman Filter (UKF) provides improved estimates when compared with the Extended Kalman Filter on the same dataset. This is mainly due to the CTRV motion model used in the Unscented Kalman Filter, which is better able to model acceleration (change in direction). The CTRV motion model contains non-linear equations which the UKF can accomodate. Comparatively the EKF used the Taylor series to create linear equations that approximated non-linear motion. Furthermore the table shows that using both lidar and radar sensors improves accuracy over using just one sensor. Therefore the UKF is an effective method for sensor fusion. ### Filter Consistency The process noise paramters were tuned based on the Normalize Innovation Squared (NIS) values produced by the filter. The two graphs below show the NIS values from the final filter implementaiton (for laser and radar respectively). The 95% line indicates that only 5% of data points should be above this line for a consistent estimate. For the radar NIS graph approximately 4.4% of values are above the line, which is close to expected. However for the laser graph only about 2% of data points are above the line, which indicates that the filter is over estimating the process uncertainty. This is something that can be tuned further in the future implementations. For the time being the filter provides RMSE values well below project specifications. ![alt text][image3] ![alt text][image4] **Note:** the 95% line is different for laser and radar since they use 2 and 3 degrees of freedom respectively.

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：