Thesis List

Event cameras are promising for mobile robot perception due to their high temporal resolution and robustness to motion blur. However, event-based detection and segmentation are strongly affected by camera ego-motion, which can create background clutter and unstable object boundaries during event accumulation. This thesis investigates how different pose sources can support event-based 2D perception. The student will compare selected event-based visual odometry / SLAM methods with a motion-capture trajectory used as ground truth. The evaluated pose sources may include MoCap, robot odometry, stereo event-based VO/SLAM, and optionally RGB-D odometry.

Using these pose estimates, the student will implement pose-assisted event representations for ego-motion compensated event accumulation. These representations will be compared against pose-agnostic event frames for downstream 2D object detection and, if feasible, foreground or object segmentation. The work will analyze how pose accuracy, accumulation window size, and motion type influence detection or segmentation quality. The thesis will provide both a quantitative evaluation of event-SLAM/VO methods against MoCap and an analysis of whether pose accuracy translates into improved event-based perception.

Requirements: 

• Solid programming skills in Python
• Basic knowledge of computer vision
• Understanding of camera models, projections, and coordinate frames
• Familiarity with robotics concepts, especially transformations and ego-motion
• Basic understanding of visual odometry or SLAM
• Interest in event-based vision
• Ability to work with experimental data and evaluate results quantitatively

Event cameras are promising for object tracking due to their high temporal resolution and robustness to motion blur. However, event cameras do not directly provide depth, and observations from a mobile robot are influenced by both object motion and camera ego-motion. This thesis investigates how event-based object detections, depth information, and camera-pose estimates can be combined to obtain stable 3D object trajectories.

The student will develop an offline 3D object-tracking pipeline using pre-recorded event-camera, RGB-D, odometry, and motion-capture data. An existing event-based object detector will provide 2D detections, which will be associated with RGB-D measurements to estimate object positions in three-dimensional space. Camera poses from MoCap, robot odometry, or visual odometry/SLAM will be used to transform these observations into a common robot or world coordinate system.

Different depth-association and tracking approaches will be evaluated, including robust depth extraction within object detections, temporal filtering, and data association. The work will analyze how pose accuracy, depth quality, event activity, camera motion, and object motion influence 3D localization and tracking performance. An optional extension is to develop a reliability-aware tracking method that handles missing depth, sparse event activity, or uncertain pose estimates.

The thesis is primarily based on pre-recorded datasets. Calibration parameters, synchronized recordings, and reference trajectories will be provided. The main focus is on algorithm development, offline evaluation, and quantitative analysis rather than hardware integration or data collection.

Requirements: 

• Solid programming skills in Python
• Basic knowledge of computer vision
• Understanding of camera models, projections, and coordinate frames
• Familiarity with robotics concepts, especially transformations and ego-motion
• Basic understanding of object detection, tracking, or state estimation
• Interest in event-based vision and multimodal perception
• Ability to work with experimental data and evaluate results quantitatively