Human behavior understanding is central to enabling intelligent, real-time, and proactive systems to interact with humans. While numerous studies have been dedicated to human behavior understanding, detecting and predicting sudden human action changes remains a formidable challenge due to the erratic and stochastic nature of human behaviors. This dissertation presents the development of efficient and accurate algorithms for understanding human sudden change behaviors and predicting such changes to facilitate intelligent systems to interact with humans seamlessly. A two-stage action change detection framework is first introduced to enable a machine to efficiently perceive and comprehend diverse human actions with variable time lengths. In the first stage, the algorithm uses a lightweight multi-window change point analysis method to identify action change times in continuous action sequences. In the second stage, an action recognition algorithm is selectively applied when a new action is detected. Numerical experiments show that this approach reduces computational costs and improves classification accuracy when compared to existing well-known action segmentation methods. Moreover, to enable robots capable of interacting with human seamlessly and proactively, it is essential to develop their cognitive ability to predict human behaviors. This problem is particularly challenging due to the stochastic nature of people’s future behaviors. Existing behavior prediction algorithms, such as trajectory prediction studies, focused on producing multi-modal feasible paths to capture prediction uncertainty. However, this approach encountered challenges in predicting abrupt motions, such as sharp turning or stopping, due to the infrequent occurrence of such scenarios in the current public datasets and the absence of explicit cues in the observed motion pattern. A novel approach is proposed to predict a potential abrupt change in a pedestrian’s motion before this change can be observed from their current motion pattern. The approach is particularly focused on a group of pedestrians undergoing a potentially changing scenario, where one of the group members may lead the change. By detecting the change and leveraging this information, the unforeseen change in another group member’s motion can be predicted. The proposed approach is validated on a novel dataset consisting of abundant abruptly changing scenarios in group settings, demonstrating notable improvement in prediction accuracy compared to existing approaches.