Automatic pain assessment is crucial in clinical situations where verbal communication is limited or impossible. Traditional pain evaluation relies heavily on subjective reporting, leading to inconsistent and often unreliable diagnoses. This paper reviews recent approaches in automatic pain detection, including facial expression analysis, physiological signal processing, thermographic imaging, and multimodal fusion. We then propose a novel method combining infrared thermography with deep convolutional neural networks (CNNs) for pain localization and intensity classification. Our system acquires thermal images, segments pain-affected regions via a thermal segmentation algorithm, and classifies pain intensity into three categories : no pain, moderate pain, and severe pain. Experimental validation on a balanced dataset demonstrates promising accuracy, sensitivity, and specificity. The proposed approach offers a non-invasive, objective, and efficient solution with potential clinical applicability. Limitations and perspectives for real-world deployment are discussed.