Abstract
Anomaly detection from video surveillance inputs helps to improve security in crowded places and outdoors. The captured image is analyzed to identify human faces, objects, and abnormal events through computer-aided analytics. This article proposes a Texture-Classification-based Feature Processing (TCFP) technique for distinguishing anomalies in captured video inputs. The anomalies are identified as events from the sequence frames wherein the dynamic inputs are distinguished using their features. Deep learning is employed for temporal training features based on frame characteristics in this distinguishing process. The input frame is segregated using textural boundaries separated using non-dimensional features. The learning process trains dimensional and non-dimensional features for identifying anomalies and maximizing detection accuracy. The textural boundaries are defined using the non-dimensional vectors present in the frame series in the different face classifications. Therefore, the errors are confined within selective boundaries without impacting the preceding feature. This improves the F1score with less processing time.
•To increase different face classification accuracy by identifying the texture boundaries using non-dimensional vectors•To minimize the processing time and violence detection accuracy using effective classifiers•To reduce the misclassification error rate using the optimized network parameter updating procedure