Driving is a complex task due to an increase in traffic density and distractions. Traffic safety and efficiency are essential and a fundamental pillar to achieve this is predicting the intention of traffic participants for improving ADAS or autonomous driving systems. This thesis performs a comparative analysis between sequence-based neural networks in turning intention recognition of cars at a roundabout in Gothenburg. The dataset used contained real-world data of car trajectories and was provided by Viscando. Due to the complexity of roundabouts, which are complex traffic scenarios, where the interaction between cars and uncertainty increases, recurrent neural networks, which have the ability of learning patterns from sequential data, were suitable for this task. The methodology combined the CRISP-DM framework with an experimental approach, which allowed for a fair and structured evaluation of the models. The performance of recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) models were compared against a logistic regression (LR) baseline. The findings showed that all sequence-based models outperformed the baseline model, achieving 99.X% of accuracy 1 second before the maneuver was done. However, the performance differences between the recurrent networks were minimal, demonstrating that in this context, an increase in architectural complexity does not necessarily imply better performance.