A novel method for detecting normal and abnormal crowd dynamics using optical flow and energy level analysis
M.N. Renukadevi, Kavyashree I Pattan, T.M. Rajesh, Praveen Kulkarni · 2025
This work presents a novel approach that combines energy parameter computations and optical flow techniques to identify anomalous crowd behavior in videos. We estimate the mobility between consecutive video frames using the Horn-Schunck and Farneback optical flow methods, paying particular attention to the direction and amplitude of the velocity field for each pixel. We also compute energy parameters to measure the textural and intensity distribution properties of video frames: contrast, entropy, and uniformity. By comparing these descriptors‘ mean and standard deviation to predetermined thresholds, they aid in the distinction between typical and aberrant crowd behavior. Using the UCF dataset, our method effectively shows that average optical flow magnitudes for normal crowd videos stay below the threshold while exceeding it for anomalous recordings. Different patterns emerge from the examination of energy parameters: aberrant videos exhibit the opposite trend, while normal videos have higher uniformity and lower entropy. Our approach outperforms the standalone Horn-Schunck and Farneback approaches in detecting aberrant crowd behavior, with a 91% accuracy rate. Evaluation criteria including precision, recall, and F1 score further verify our method. This offers a dependable means of improving security and safety in public areas through efficient crowd behavior monitoring.