This paper presents a systematic review of models, datasets, and evaluation metrics used for pedestrian detection, trajectory forecasting, and intention prediction in urban pedestrian–vehicle mixed environments. As autonomous driving systems increasingly interact with vulnerable road users (VRUs), accurately anticipating human motion and intent becomes essential for safety, trust, and reliable decision-making. We employed a PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology to gather the latest research in this field, conducting a systematic search across IEEE Xplore, ScienceDirect, and Google Scholar. From an initial set of over 350 records, we screened studies based on predefined inclusion criteria and retained 60 primary research works for detailed qualitative synthesis. The review develops a unified taxonomy spanning perception, short- and long-term trajectory prediction, and intention recognition, while comparing traditional models, deep learning architectures, generative frameworks, diffusion models, reinforcement learning, and emerging large vision–language models. We further analyze widely used datasets (e.g., ETH/UCY, SDD, PIE, JAAD) and standard evaluation metrics such as ADE (Average Displacement Error) /FDE (Final Displacement Error), alongside risk-aware and intention-specific measures. We identify limitations in current approaches, generalization gaps, and limited explainability and outline key future research directions for developing human-centered predictive systems capable of robust performance in complex urban traffic scenes.