Contributions to tracking and artificial intelligence based Lidar Signal Processing for automotive applications : point cloud processing and augmentation for AI,optimized vehicle and lane marker tracking for automotive / von M.Sc. Martin Alsfasser. Wuppertal, 2022
Inhalt
- Abstract
- Contents
- Contents
- 1 Introduction
- 2 Fundamentals
- 3 Advancements in Small Object Tracking in Camera Images
- 3.1 Challenges of Automatic Headlight Control
- 3.2 Gaussian-Mixture Probability Hypothesis Density Filter for Multiple Target-Multiple Detection Tracking
- 3.2.1 Definition of a Gaussian-Mixture Probability Hypothesis Density Filter
- 3.2.2 Optimizing the GM-PHD for Vehicle Light Tracking
- 3.2.3 Finding an optimized parameter set
- 3.3 Solving Asynchronism Between PWM Light Sources and CMOS Image Sensors
- 3.3.1 Lucas-Kanade Optical Tracker
- 3.3.2 Regularization Methods to Stabilize KLT Tracking in Noisy Environments
- 3.3.3 Achieving Higher Precision Optical Tracking due to Pyramidal Evaluation
- 3.3.4 Forward-Backward Verification of Tracking Results
- 3.4 Combinatorial use of Optical Tracking and Track Prediction for Stable Object Tracks in Adverse Conditions
- 3.5 Modelling Car Column Movement by Swarm Movement
- 3.5.1 Advantages of Group Tracking for Vehicle Light Tracking
- 3.5.2 OPTICS for Clustering Light Sources
- 3.5.3 Virtual Leader-Follower vs. Cucker-Smale Flocking Model
- 3.5.4 Predicting Group Movement
- 3.6 Evaluation of Advanced Light Source Tracking Components
- 4 Improving Lidar Object Detection Algorithms
- 4.1 A Summary of State of the Art Lidar Object Detection Algorithms
- 4.2 Exploring Structured Approaches to Point Cloud Processing
- 4.2.1 Issues and Options With the Classical Square Grid Based Approach
- 4.2.2 Advantages & Disadvantages of Sphere Based Grids
- 4.3 Improving Detection Results by Adding Hand-Crafted Features
- 4.3.1 Encoding Height Information in 2D Feature Map
- 4.3.2 Occupancy Grid Maps as Additional Feature Layer
- 4.4 Novel Methods of Input Data Augmentation for Improved Network Generalization
- 4.5 Training, Network Structure and Challenges
- 4.6 Quantifying Runtime and Memory Advantages and Evaluating Detection Results
- 5 Innovative Semiautomatic Data Annotation Methods for Enhanced Annotation Efficiency
- 5.1 Improving Training Data Generation With Neural Network Support
- 5.2 Baseline Network Structure for Offline Annotation Algorithm
- 5.2.1 Data Preparation and Pre-Processing
- 5.2.2 Feature Extraction Structure
- 5.2.3 Class and Bounding Box Regression
- 5.3 Advances in Network Structure
- 5.3.1 Time Series Considerations for Improved Result Stability
- 5.3.2 Improving Network Performance by Patch-Wise Data Processing
- 5.3.3 Structure Aware Point Cloud Augmentation
- 5.3.4 Clustering Results as Input Feature
- 5.4 Evaluation of Network Improvements Against Baseline
- 5.5 Training and Performance Evaluation of SAPCA vs Angular vs. Naive Injection
- 6 Lane Marker Detection with Lidar + RGB camera sensor fusion
- 6.1 Introduction to Lane Marker Detection in Lidar Point Clouds
- 6.1.1 Preparing the Point Cloud for Further Processing
- 6.1.2 Lane Marker Detection
- 6.1.3 Refining a Spline Model to Track Lane Markings
- 6.1.4 Evaluation of Lane Marker Results
- 6.2 Projection and Reprojection for RGB + Lidar Fusion
- 7 Conclusion and Outlook
- 7.1 Conclusion and Evolution from Head- and Taillight Tracking
- 7.2 Advancing Fast Object Detection in Lidar Point Clouds
- 7.3 Applying Complex Neural Networks to Data Annotation
- 7.4 A Classical Approach to Lane Marker Detection in Point Clouds
- Bibliography
- List of Figures
- List of Tables
- List of Listings
- Acronyms
- A Appendix