The master’s thesis extends the FIVIS bicycle simulator, originally developed by HHK+08 and later updated in 2022, by integrating a temporal reasoning module that allows non‑player characters (NPCs) to infer new knowledge from previously acquired information. The added rule‑based inference system uses tags and a state‑history mechanism to enable motorised NPCs to anticipate pedestrian intentions early and incorporate this anticipation into their decision‑making process. The inference engine evaluates pedestrian intention tags such as APPROACH_PEDESTRIAN_CROSSING, ON_AGENT_LANE, and LEAVE_PEDESTRIAN_CROSSING, applying confidence thresholds defined in a PedestrianIntentionProfile. For example, a confidence of 0.3 for approaching a crossing, 0.0 for staying on the lane, and 0.7 for leaving the crossing are used in the standard profile, while modified profiles adjust these values to 0.5, 0.0, and 0.8 or 0.74, 0.1 in other scenarios. These thresholds are combined with Markov‑chain transition probabilities to model the stochastic evolution of pedestrian states.

Three realistic traffic scenarios were implemented to evaluate the system. In scenario 1.1, an agent with a target speed of 13.89 m s⁻¹ starts from position A1, while a pedestrian begins at F1, 7.96 m from the crossing centre. Scenario 1.2 uses the same agent speed but starts from A4, with the pedestrian at F3, 20.22 m away. Scenario 2.1 introduces a slower agent speed of 8.33 m s⁻¹ from A3, with the pedestrian again at F3. In all cases, the standard rule set and standard Markov transition probabilities are applied. Scenario 3 explores abrupt reversal of pedestrian motion; four variants modify either the rule set or the transition probabilities, demonstrating how changes in inference thresholds influence the likelihood of collision events. The simulation parameters show that rare accident scenarios between virtual pedestrians and drivers emerge only when the decision process is influenced by incomplete, error‑prone, or ambiguous input data, thereby reproducing human‑like uncertainty and error in NPC behaviour.

Performance results indicate that the extended simulator can generate plausible, non‑deterministic accident scenarios without compromising real‑time simulation speed. The agent’s lane‑change decisions are evaluated within milliseconds, and the inference engine’s tag evaluation loop processes all perceivable pedestrians in a single frame, maintaining the simulator’s frame rate above 60 fps on a standard workstation. The use of Markov‑chain probabilities allows the system to produce statistically realistic pedestrian trajectories, with transition matrices such as (0.3, 0.0, 0.7) for the standard profile and (0.5, 0.0, 0.8) for modified profiles, ensuring that pedestrian behaviour varies across simulation runs.

The project was carried out as part of a master’s programme at a German university, supervised by a faculty member in the computer science department. No external partners were involved, and the work was funded by the university’s internal research grant allocated to the student’s thesis project. The timeframe covered the academic year 2023–2024, during which the author developed the inference module, implemented the three scenarios, and performed extensive validation against expected pedestrian behaviour patterns. The collaboration remained strictly within the university, with the supervisor providing guidance on the integration of the rule‑based system into the existing FIVIS architecture and on the design of the evaluation scenarios.