Qixuan Zhang

Wearable Camera based Pose Estimation of Driver's Head: VO/VIO System Development; Monocular Camera based V-SLAM implementation. Apriltag Detection; Keypoints inside apriltags extraction and correction with prior; Keypoints tracking with TAPIR and Optical Flow; Muiti-Sensor System Fusion and Calibration with ROS(Camera, IMU); Camera Parameter Control (Exposure, gain); Camera Pose Estimation via Pnp and RANSAC and Global Optimization with Bundle Adjustment.

Hybrid Maching Learning for Articulation Angle Estimation of Truck Trailer Develpment of algorithms based on images and kinematic vector inputs: Image data preprocessing, Data augmentation. CNN-based Model training, Transformer-based Multi-Modal Model deveploment and training. Employ Extended Kalman Filter to fuse kinematic model and machine learning model. Dynamically adjust the observation covariance based on epistemic uncertainty of machine learning models. Domain Adaption: Out of Domain Detection and uncertainty estimation, Supervised Contrastive Learning Approach validation. Test and Evaluation: Test and evaluate developed algorithms on different challenging out of domain data sets(different type(appearance) of trailer, low Illumination condition). Post-hoc explainability validation of machine learning models: Modificate Grad-CAM approach to fit regression task, feature heat map visualization.

Secnario simulation in CARLA: Simulate different LiDAR deployments on truck-trailer combination, visualize opints cloud, analysis blind spots of LiDAR perception for Ego-truck-trailer in different scenarios. Simulate test scenarios from Euro-NCAP in CARLA, program the behaviors of every road user with python. Automatically anonymize human faces and licences plates: Employ YOLO v8 to detect human face and licences plates, track these objects with DeepSORT in frames where target objects are failed to be detected. Add blur filter on bounding boxes.

Daten Annotation und Visualisierung