FPI-AD: Anomaly Detection with Dense Skeleton Analysis and Future Frame Prediction

In this study, we introduce an innovative approach for recognizing abnormal behavior by leveraging predictive models to forecast future frames in video sequences. Our method focuses on pose images that include only the human dense skeleton, addressing privacy concerns more effectively than traditional techniques that rely on RGB images. To detect abnormalities, we generate an abnormality score by analyzing the differences not only between the predicted future frame generated by our model and its groundtruth but also between the humans present in these frames. This dual analysis allows us to classify a predicted future frame as either abnormal or normal by comparing it with the actual frame in the sequence. Significant deviations between the predicted and actual frames are indicative of abnormal events, as the model, trained predominantly on normal behavior, tends to struggle with accurately predicting rare or atypical occurrences. Additionally, we introduce a new dataset NRIPS which includes a wider range of ab-normal actions occurring in more complex and multi-object scenes. Experiments are conducted in both the public dataset SanghaiTech and our proposed dataset NRIPS, yielding AUC results of approximately 68% and 78% respectively, demonstrating the superiority of our method over previous methods.