Riassunto analitico
Road safety remains a critical concern in nowadays urban environments, especially regar-
ding Vulnerable Road Users (VRU) such as pedestrians. To mitigate potential accidents,
Intelligent Transport Systems (ITS) have emerged as promising solutions, leveraging ad-
vanced technologies to enhance road safety. In this thesis work, we focus on pedestrian
trajectory forecasting, a crucial issue to solve, which allows to anticipate the movements
of pedestrians and enable proactive safety measures.
This study proposes the implementation of a transformer network model for pedestrian
trajectory forecasting. Transformers have shown great capabilities in modeling sequential
data and they are widely used in the research field of Natural Language Processing (NLP)
research field. By leveraging their attention mechanisms and ability to capture long-term
dependencies, they have recently shown good results also in forecasting trajectories.
The primary objective of this thesis is to compare and evaluate three different training
approaches for the proposed transformer network model: local, centralized, and federated
training. In the local training approach, data acquired by individual sources, such as vehi-
cles equipped with cameras and sensors, are used to train a local model. The centralized
training approach involves collecting all the data from various vehicles at a central server,
where the model is trained using the combined datasets. Finally, the federated training
approach aims to maintain data privacy and security by training the model collaboratively
across multiple decentralized sources, without sharing the raw data.
We will assess the performance of each training approach by considering aspects such
as accuracy, and training time. Various real-world pedestrian trajectory datasets will be
utilized to validate the proposed models and compare the training approaches. Additio-
nally, the Mean Average Displacement (MAD) and Final Average Displacement (FAD)
metrics will be employed to quantify the prediction accuracy and the performance the
trained models achieve.
The findings of this research work are expected to provide some initial valuable in-
sights into the suitability and effectiveness of different training approaches for pedestrian
trajectory forecasting within the ITS context. The results will contribute to enhancing
road safety measures, enabling proactive interventions, and reducing the risk of accidents
involving VRUs.
|
