Determining human intention in videos 1

• Main Author: Tan, Rachel
• Other Authors: Cham Tat Jen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166016
• Institution: Nanyang Technological University

With the recent heightened interest in autonomous vehicles in consumer markets, there are many attempts to reduce the accident rates for both semi-autonomous and fully-autonomous vehicles. One such attempt is in developing the ability of the vehicle to accurately predict the future trajectory of pedestrians and other vehicles within a short period. To address this problem, this paper studied the Trajectron++ model, a machine learning model which has outperformed other popular and traditional machine learning models in trajectory prediction. In an attempt to improve on the Trajectron++ model, the Long-Short-Term Memory (LSTM) Network is replaced with Transformers as Transformers are found to outperform LSTMs in other similar problems such as Natural Language Processing. Evaluation results proved that the replacement of LSTMs with Transformers result in a less accurate prediction than the original Trajectron++ model. This is deduced from the existence of a larger maximum value for the Average Displacement Error and Final Displacement Error in most datasets and a generally higher Kernel Displacement Error Negative Log-Likelihood (KDE NLL) value for the Trajectron++ with Transformers model. Notably, there was a contradiction observed in the KDE NLL values for different datasets which suggest that the modified Trajectron++ model does perform better for one particular dataset. To explain what caused this contradiction and explore possible areas of improvement to the modified Trajectron++ model, future research could be performed.