We are constantly looking for new students of local universities to join us to work on our autonomous cars. Background in Robotics (i.e. Robotics class or relevant experience with ROS) is a requirement for the autonomous cars related topics. For possible open thesis topics please contact Daniel Goehring.
Here is a non-complete list of topics which might be of interest for you:
Camera based:
Multi-Task Learning using Bayesian Deep Learning uncertainty for Scene Geometry and Semantics (WM-RC).- Deep Neural Networks (CNN) for trajectory prediction based on images and odometry data
- Deep Neural Networks (CNN) for Swarm behavior learning, e.g., learning the desired position of a car within a group of vehicles
Fish-eye camera depth estimationFish-eye camera soiling detectionFish-eye camera traffic sign detection and classificationFish-eye camera weather condition classification- Passenger on- and offboarding detection and tracking
Road damage detection and classification (RDD2022 Dataset) and 3D localization- Road pictogram detection and classification
Traffic light recognition using 2d or 3d camera dataTurn indicator (blinker) recognition using 2d or 3d camera data- Visual slam in traffic situations – using visual features and/or optical flow
- Dataset Generation using Stable Diffusion and ControlNet
- Guided dataset generation using Segment Anything
- DepthAnything and ZoeDepth for realtime depth estimation
LiDAR based:
LiDAR construction side beacon detection- LiDAR IMU-Coupling (use the accelerometer to avoid LiDAR measurement errors due to vibrations)
- LiDAR obstruction mapping (would include ground segmentation and 3d
occluded space modeling) - Self-localization and mapping using LiDAR data or point cloud data from stereo cameras
- Sensor fusion of data from different sensors, e.g., LiDAR with camera data
- Classification of LiDAR objects using image based classification and the projection of 3d-points into the 2d image
Virtual Velodyne scanner from Berlin’s pointcloud data- Dynamic obstacle simulation using a virtual LiDAR sensor
- Semantic Segmentation (VLS 128 / HDL-64)
- Panoptic Segmentation (VLS 128 / HDL-64)
- Obstacle detection and tracking (VLS 128 / HDL-64)
- LiDAR inertial odometry
- LiDAR wall and curb detection
- Cross modality localization, e.g. visual localization in LiDAR maps
- 3D scene reconstruction using LiDAR and/or camera data
- LiDAR cooperative perception
- Deep Neural Network-based LiDAR data compression methods for constrained communications
IMU based:
- Automatic speed bumper detection using smartphones
V2X-Communication:
- Cooperative driving maneuvers: Forming an emergency corridor
- Cooperative driving maneuvers: Allow bus to merge into traffic
- Cooperative moving object tracking
Planning:
- Calculation of drive splines from road boundary and lane line map data
- Deep-Learning based trajectory planning, Deep Reinforcement Learning
- World modeling, planning within a structured dynamic environment
- Planning of lane changes, merging into traffic using timed elastic bands, or A*, Rapidly Exploring Random Trees
- Obstacle prediction based on map data
Optimal control:
- Developing different control strategies to drive on a map and with dynamic objects
- Analysis of dynamic properties for swarm based control strategies, given certain uncertainties
Simulation:
Virtual reality obstacle simulation
Autonomous Modelcar Project Topics:
- Lanelet2 force field navigation including dynamic obstacles
- Model predictive control
- Traffic Light detection and tracking
- Obstacles detection and tracking
- Autocharging using inductive charging