This paper introduces a comprehensive approach to crowd anomaly detection by dynamically switching between macroscopic and microscopic levels based on the crowd density. With the growing requirement of automated systems to detect anomalies in crowds, our method aims to avoid crowd-related disasters such as human stampedes, overcrowding and riots. Advanced multiple object tracking techniques are implemented with graph-based MOT, RAFT and LSTM to monitor the crowd behavior. This integrated approach improves our ability to detect potential threats in real time, making public spaces safer. In case of dense crowd, we use macroscopic analysis, wherein we use flow fields and other video features, to train a LSTM model that detects any abrupt changes and returns an anomaly probability. For microscopic analysis, we capture crowd movements using YOLO and a sudden change in movement is considered an anomaly. Our proposed model achieves accuracy of 95% on UMN and 90.47% synthetic dataset.