Traffic safety in automated vehicle (AV) research focuses on ensuring safe interactions with other road users. The key challenge lies in understanding and predicting vulnerable road users (VRUs) behavior in traffic, with cyclist-related research remaining relatively limited compared to pedestrians. To address this research gap, this paper proposes a two-stage deep learning model for predicting cyclists’ maneuvers at an unsignalized intersection. The model is based on Bidirectional Long Short-Term Memory network (B-LSTM) and predicts tactical maneuvers (left turn, right turn, or straight crossing) before cyclists enter the intersection. Data is collected using a bicycle simulator that enables interaction with simulated road users in a virtual reality (VR) environment. The cyclist video data serves as the only input to the model, eliminating the need for external devices. The first stage consists of two parallel models that classify cyclists’ explicit and implicit gestures, which serve as communication signals. In the second stage, these classified gestures are used to predict future tactical maneuvers, with predefined gesture weightings assigned based on their correlation with maneuver categories. Results demonstrate the importance of incorporating cyclists’ communication signals, especially implicit gestures, and predefined gesture weightings over model development in prediction accuracy and robustness.