How Artificial Intelligence Learns to Perceive Traffic Scenes Beyond Its Training Data: Reliable traffic perception under changing environments through self-training, teacher–student learning, and feature alignment.

Unsupervised domain adaptation (UDA) is critical for generalizing traffic scene understanding across diverse environments where manual data annotation is not feasible. Modern object detection models frequently fail during deployment due to domain shift, in which detection performance drops due to variables such as weather, lighting, and urban architecture. This paper examines strategies to bridge this gap by encouraging models to learn from unlabeled target data through self-training. In this approach, the model generates pseudo-labels from its own high-confidence predictions to supervise optimization. To ensure stability in this process, we analyze the teacher-student framework, which defines a slow-updating teacher model to provide reliable guidance and mitigate error reinforcement due to noisy pseudo-labels. Furthermore, we explore feature alignment methods, such as adversarial, contrastive, and optimal transport methods, to learn domain-invariant semantics over superficial appearance cues. Through the integration of these methods, a framework is developed to build robust traffic scene understanding systems with high accuracy in diverse environments.