Driving Control as a Source of Variation in Visual Motion and Pedestrian Tracking

Kurzfassung

Simulation-based testing allows automated-driving scenarios to be repeated under controlled conditions. However, identical nominal scenario definitions do not necessarily lead to identical recorded data. The realized ego-vehicle trajectory, speed profile, camera viewpoint, and interaction timing can differ depending on how the ego vehicle is controlled. This thesis investigates whether human-operated and BehaviorAgent-controlled executions of matched CARLA scenarios lead to measurable differences in visual motion structure and pedestrian-tracking performance. Three urban CARLA scenarios with turning maneuvers and pedestrian crossings were executed repeatedly under both ego-control modes. The scenario definition, including map, route, pedestrian triggers, weather, and sensor setup, was kept fixed, while the realized executions were allowed to differ between modes. Incomplete or invalid repeats were excluded before analysis. The evaluation uses two complementary branches: repeat-level optical-flow descriptors from simulator-provided CARLA optical flow and RGB-derived Farnebäck flow, and pedestrian-tracking metrics computed for multiple tracking-by-detection methods using both oracle and YOLO11x-based detections. The results show that ego-control mode affects both visual motion structure and pedestrian-tracking conditions, although the effect depends on the scenario. Human- operated repeats generally show higher mean flow magnitude and stronger between repeat variability. BehaviorAgent-controlled repeats are more repeatable in several global motion descriptors, but their local roughness descriptors do not consistently indicate smoother image motion. Tracking results also differ between modes. In the less complex scenarios, BehaviorAgent-controlled repeats often achieve higher visible IDF1 and fewer visible identity switches. In the most complex double-crossing scenario, the mode effect is weaker and more dependent on the tracker and detection source. The combined analysis indicates that selected optical-flow descriptors are associated with tracking difficulty, with visible identity switches tending to occur in windows with stronger or less stable image-plane motion. Overall, the thesis shows that ego-control mode should be treated as an experimental factor in simulation-based perception evaluation. The findings are specific to the investigated CARLA setup, scenario set, sensors, trackers, detection sources, and quality-filtered repeat set.