Powered Two-Wheeler (PTW)s are a class of Vulnerable Road Users (VRU). They are easily susceptible to accidents due to the inherent and unique kinematics of PTWs that have different dynamics while in motion and while stationary. Roll angles of PTW are highly correlated with their lateral dynamics and complement existing safety systems like the Advanced Rider Assistance System (ARAS) and Motorcycle Stability Control (MSC). This thesis work aims to address the safety concerns by contributing to roll angle predictions in a multi-modal fashion using a data-driven approach. Unlike most state-of-the-art research on Multi-Modal Trajectory Prediction (MTP), which used multiple information modalities, the proposed research work centered around using vehicle dynamics data along with a few rider behavior signals. The constraint on the available modalities of data posed a significant challenge, and the proposed method overcomes these challenges by using deep representation learning techniques on the available dataset. Initially, an Autoencoder (AE) was used to reduce the dimensionality of the input data to enhance clustering. To introduce the notion of modality, Gaussian Mixture Model (GMM) was used over the condensed data to identify clusters in an unsupervised approach. The identified clusters were then used to formulate modes, the cornerstones for envisioning a multi-modal approach. The distinct clusters encapsulate data points that exhibit different variations in vehicle dynamics. It was found that clusters with high variations were associated with high lateral dynamic activities such as lane changes, lane merges, and turning. They were distinct from groups exhibiting low dynamics, such as straight riding. The proposed MTP hybrid model, realized using the identified clusters, consists of two significant parts: the Long Short-Term Memory (LSTM) classifier that predicts the future modes and a set of LSTM units, each dedicated to giving roll angle predictions belonging to a particular mode. The MTP model was tested under ideal and realistic conditions. While we have seen promising improvement in roll angle predictions in the ideal situation, there are challenges when predicting outcomes with a noisy classifier. The significant class imbalance in vehicle trajectory information makes it hard to explore ways to add more data without the risk of mislabelling. Variational Auto Encoder (VAE) and other data augmentation techniques help mitigate class imbalance. However, in the context of PTW trajectory data, additional overhead is needed to validate if the generated data aligns well with the unique physics governing the stability of PTWs. The roll angle outcomes tested under the two conditions showed that with limited data modality, MTPs were feasible, but accurate information about future conditions was crucial for improving prediction accuracy. Our findings set the stage for future improvements in PTW safety systems, emphasizing the importance of having precise information about future conditions.