Crowd anomaly detection and localization via deep convolutional model with improved spatio temporal textures.

Detecting anomalies in crowd scenes holds a critical task in automated video inspection to prevent any casualties in the regions that witness the higher quantity of footfalls. The main confront for the automated classification of crowd anomalies in images is the utilization of feature sets and methods that could be simulated in each crowded scenario. This research aims to provide a novel crowd anomaly detection approach with 4 key steps. Initially, video frame conversion takes place during preprocessing and then the segmentation process is done via Entropy based FCM (EFCM). Further, a Histogram of Gradients (HoG), Local Gradient Pattern (LGP), and improved spatiotemporal features are derived. The DCNN classifier is then used to classify these traits, and the exact discovered results are obtained. For exact detection, this paper intends to optimize the weights of DCNN using Self Improved Dingo Optimization (SI-DOX) model. The proposed technique has achieved high accuracy values at the 90th LP, better than the existing methods like DCNN + DOX, DCNN + SNFO, DCNN + PRO, DCNN + GOA, and DCNN + GWO by 3.23%, 10.75%, 4.3%, 3.23%, and 4.3%, respectively. Finally, the performance of proposed model is evaluated regarding diverse metrics. [ABSTRACT FROM AUTHOR]