Human activities that deviate from the norm are deemed abnormal, and such individuals are referred to as anomalous objects. Employing visual data to detect abnormal behaviour is a complex topic in video processing. This research proposes a novel method for detecting abnormal behaviour in complicated, crowded environments. In this article, we proposed a robust method for abnormal action recognition. We initially processed the data, applying fuzzy c mean and super pixel-based segmentation, extracting the features and tracking the object. The next step is to optimize the data. We used a deep data mining approach via t-distributed stochastic neighbor embedding procedure, and for classification, we applied random forest. We achieved 80.24% accuracy rate for human detection over UCSD dataset, and 79.19% for Shanghai tech dataset. We also got 84.00% accuracy of abnormal action recognition over UCSD dataset and 82.00% over Shanghai tech dataset.