Emotion detection and facial expression recognition have emerged as critical aspects of human–computer interaction (HCI) as technology strives to understand human feelings in different fields such as health, security, sales, teaching, and entertainment. The development of the ML and DL technologies in recent years has greatly enhanced the performance of the emotion recognition systems in terms of accuracy, scalability, and flexibility. Facial expressions, vocal intonations, and even body posture and gestures are now distinguishable to identify emotions as the data is processed in real time. However, there are some issues that are still unresolved even as the field advances, including data heterogeneity across demographics, model evaluation in real-world contexts, and the latency issues imperative in real-time applications. Moreover, demographic bias in the emotion recognition system and ethical concerns over data privacy and fairness are emerging concerns which should be addressed. This paper provides an updated survey on FER, covering the traditional approaches, the DL approaches, the fusion of both paradigms, and the performance assessment methodologies. In the same way, potential limitations of current FER systems are discussed as well as possible directions for future research, such as improvements of FER models and their robustness, augmenting the variety of datasets used or designing countermeasures against biases.