<p>In video surveillance, pedestrians’ gaze toward a certain area reflects their potential intentions and perceptual states, providing crucial clues for event analysis. Taking gaze points as the core information carrier, we construct a dataset with bus stops as the scenario and propose an event analysis approach based on variations in crowd gaze points. The method uses Vision Transformer to extract global features, fuses global and local information via a U-shaped network to improve gaze point regression accuracy. To tackle the "weak gaze point" phenomenon characterized by heavy pedestrian flow, scattered gazes and lack of focused areas in the scene, we introduce a saliency detection network, which enhances visual features of key targets while suppressing redundant background interference. For fine-grained event analysis, a semantically enhanced feature vector is built based on accurate gaze point localization. This vector integrates multi-dimensional attributes including bus stop structure, pedestrian head positions and salient targets to enable precise event analysis. To verify the generalization and robustness of the model, experiments are conducted on a self-built dataset containing multiple gaze targets and a public dataset containing a single gaze target. Results show the gaze point regression method achieves area under the curve values of 84.6% and 92.4% in two different scenarios, laying a solid foundation for subsequent event analysis. In the event analysis task, the method performs excellently with accuracy, precision, recall and F1-score reaching 98.4%, 98.5%, 98.4% and 98.4% respectively, fully demonstrating its good event analysis capability and accuracy.</p>