Simon Pointner
Cycle Safely
[thesis]

Information

Abstract

This thesis presents a design and implementation of an end-to-end collision prediction and detection pipeline tailored to cyclists. The primary goal is to predict potential collisions between the cyclist (ego agent) and other road users, particularly motorised vehicles, which pose a higher risk due to higher momentum and speed. The proposed pipeline integrates conventional techniques for solving the sub-tasks of object detection, object tracking, and trajectory forecasting. Specifically, Super Fast and Accurate 3D Object Detection (SFA3D) is used for the detection, a Kalman filter-based multi-object tracker for temporal association of these detections, and a machine learning-based model (prediction conditioned on goals in visual multi-agent settings) is adapted and trained for the prediction of future trajectories. CARLA driving simulator facilitates the training and development of the prediction model by creating a synthetic dataset of cycling and the interaction with other road users. The system is evaluated on the synthetic dataset and also on the real-world KITTI dataset, and additional ablation studies examine the contribution of each pipeline stage. Experiments demonstrate that the proposed approach is capable of achieving reliable performance in object detection and tracking tasks. This confirms the feasibility of such a pipeline under limited sensing capabilities, such as LIDAR and GPS/IMU measurements. However, trajectory prediction remains a difficult and computationally expensive task, primarily due to the lack of documented and easily deployable open-source models. The implementation comes with a visualisation framework, built from the Rerun tool, for interactive inspection of the pipeline’s intermediate and final results.The contribution of this thesis can be summarised by the implementation of a framework for cyclist collision detection. It offers insights into how conventional and machine learning methods can be combined into a pipeline, and key limitations and points of future work for the creation of better trajectory prediction in adverse traffic contexts are outlined.

Additional Files and Images

thesis
thesis

Weblinks

BibTeX

@mastersthesis{pointner_simon-2022-maa,
  title =      "Cycle Safely",
  author =     "Simon Pointner",
  year =       "2025",
  abstract =   "This thesis presents a design and implementation of an
               end-to-end collision prediction and detection pipeline
               tailored to cyclists. The primary goal is to predict
               potential collisions between the cyclist (ego agent) and
               other road users, particularly motorised vehicles, which
               pose a higher risk due to higher momentum and speed. The
               proposed pipeline integrates conventional techniques for
               solving the sub-tasks of object detection, object tracking,
               and trajectory forecasting. Specifically, Super Fast and
               Accurate 3D Object Detection (SFA3D) is used for the
               detection, a Kalman filter-based multi-object tracker for
               temporal association of these detections, and a machine
               learning-based model (prediction conditioned on goals in
               visual multi-agent settings) is adapted and trained for the
               prediction of future trajectories. CARLA driving simulator
               facilitates the training and development of the prediction
               model by creating a synthetic dataset of cycling and the
               interaction with other road users. The system is evaluated
               on the synthetic dataset and also on the real-world KITTI
               dataset, and additional ablation studies examine the
               contribution of each pipeline stage. Experiments demonstrate
               that the proposed approach is capable of achieving reliable
               performance in object detection and tracking tasks. This
               confirms the feasibility of such a pipeline under limited
               sensing capabilities, such as LIDAR and GPS/IMU
               measurements. However, trajectory prediction remains a
               difficult and computationally expensive task, primarily due
               to the lack of documented and easily deployable open-source
               models. The implementation comes with a visualisation
               framework, built from the Rerun tool, for interactive
               inspection of the pipeline’s intermediate and final
               results.The contribution of this thesis can be summarised by
               the implementation of a framework for cyclist collision
               detection. It offers insights into how conventional and
               machine learning methods can be combined into a pipeline,
               and key limitations and points of future work for the
               creation of better trajectory prediction in adverse traffic
               contexts are outlined.",
  pages =      "86",
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Research Unit of Computer Graphics, Institute of Visual
               Computing and Human-Centered Technology, Faculty of
               Informatics, TU Wien",
  keywords =   "cycling, collision detection, trajectory prediction, machine
               learning",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2025/pointner_simon-2022-maa/",
}