Prediction of Vehicle Behavior in Urban Environments Using Machine Learning

Abstract

Autonomous driving is the primary objective for many car manufacturers today. In a fully autonomous system, a vehicle must navigate without human input, which requires a comprehensive understanding of its environment, precise control over the vehicle, and the ability to travel from one location to another. A key challenge in highway driving, particularly on multi-lane roads, is anticipating the lane-changing behavior of nearby vehicles. While turn signals are designed to indicate such maneuvers, drivers do not always use them reliably. As a result, autonomous vehicles need to detect and predict lane changes based on other cues, independent of visual signals.

This thesis addresses the problem of predicting imminent lane changes using a machine learning model based on Long Short-Term Memory (LSTM) networks. The model classifies fixed-length sliding windows into one of three categories: LLC (Left Lane Change), RLC (Right Lane Change), or NLC (No Lane Change). The work introduces two key novelties not previously explored in the literature. First, it employs kinematic features for maneuver prediction rather than relying on the vehicle’s position in a local Cartesian system or the driver’s control inputs, which are generally unavailable in real-world scenarios without V2V communication. Second, it addresses the lack of publicly available datasets by generating synthetic data using the open-source CARLA Simulator, which are then used to train the model. Finally, the model is evaluated on (a) synthetic data produced under varying conditions, and (b) real-world datasets, demonstrating its capability to generalize across different environments.