cyscgzx33 / udacitysensorfusion

Udacity Sensor Fusion Nanodegree Program

UdacitySensorFusion

Course metrials and projects of Udacity Sensor Fusion Nanodegree Program.

Lidar Obstacle Detection

Lidar Point Clouds Input	3D Tracked Objects w/ Bounding Boxes

• Contained in the folder lidar_obstacle_detection
• Implemented a 3D lidar object detection, which contains several classical algoirthm:
  • Raw Lidar Point Clouds rendering: check the top left gif
  • Processed result rendering with bounding boxes: check the top right gif
  • Lidar point clouds segmentation using the RANSAC technique (Wiki page)
  • Lidar point clouds clustering using kd-tree algorithm (Wiki page)

Type	RANSAC 2D Line Fitting	RANSAC 3D Plane Fitting	Kd-tree Clustering
Result
Source code	RANSAC 2D source code	RANSAC 3D source code	Kd-tree source code

---

Camera Object Detection

Items	Images
TTC Calculation based on 3D Object Detection
2D Image Keypoints Detection
3D Object Detection via YOLOv3

• 2D Feature Tracking
  • Implemented classic image feature detection, description as well as matching within OpenCV:
    • Keypoints detectors: implemented based on intensity gradients such as HARRIS, SHITOMASI, etc, as well as Non-maximum Suppression (NMP) technique for clearing overlapping of keypoints
    • Descriptors: applied OpenCV built-in descriptors, including Histograms of Oriented (HoG) based descriptors such as SIFT and SURF, as well as Binary Descriptors such as BRISK, BRISK, ORB and AKAZE.
    • Descriotpr Matching: implemented manually about L1, L2 norms matching, as well as K Nearest Neighbor matching algorithm based on distances / ratios
  • Performed analysis on different combinations of detector / descriptor / matching to evaluate overall performance
  • Refer to 2D Feature Tracking Page for further details
• 3D Object Tracking
  • Applied YOLOv3 object detection with trained model and generated bounding boxes for detected objects
  • Implemented 3D lidar point clouds projection on 2D camera image
  • Calculated Time-to-Collision (TTC) based on 2D camera image keypoints matching, and projected 3D lidar point clouds on 2D images
  • Conducted analysis of performance of camera TTC via various combinations of detector / descriptor
  • Refer to 3D Object Tracking Page for further details

Camera based TTC	Lidar based TTC

---

Radar Target Generation and Detection

Source / Input (from 2D FFT)	Result / Output of 2D CFAR

• 2D CFAR Algorithm
  • Refer to Radar Target Generation and Detection for further details.
  • The 2D CFAR algorithm takes the 2D FFT result, i.e., the complete Range Dopper Map (the variable RDM in the script), as its input
  • then the algorithm applies sliding window through the input, and during each iteration it conducts averaging of surrounding cell values of the interested cell (i.e., the Cell Under Test (CUT)), to take as threshold of the CUT.
  • Parameter Selection: to achieve an ideal performance of 2D CFAR, the following paramters are set:

Tr	Td	Gr	Gd	SNR
12	14	6	8	5

---

UKF Highway Object Tracking

• UKF Implementation
  • Refer to Uncented Kalman Filter Traffic Flow Tracking for further details