Accuracy Optimization with Weighted Ensemble for Multi-class Anomaly Recognition in Surveillance Videos
Preeti Sharma, Mandlem Gangadharappa · 2022
Nowadays, due to the ubiquitous use of surveillance systems it is intractable for humans to analyze large data. In video systems, anomalies occur only for a short duration of time. Hence manual monitoring results decrease in reliability and accuracy. In this work, we aim to find and combine models to detect multiclass abnormalities, so that their various parameters can contribute to the analysis. We propose MCPME-AD (Multiclass prediction model ensemble for anomaly detection). It is a model-centered ensemble that trains three predictive convolution neural networks with different parameters and combines their accuracy scores with a weighted vote classification approach and determines the best-weighted function through grid search. This method improves the detection performance. The major findings of this study are twofold. Firstly, we use ensemble-weighted learning to optimize the impact of each base model. Secondly, most studies deal with binary classification, but the reported studies do not deal with other abnormal events such as abuse, Stalinism, road accidents, robbery, etc. Our approach provides a framework for recognizing different anomalous activities using a weighted ensemble approach. This method is tested on a multi-class UCF crime video dataset and experimental results achieve higher accuracy scores than other advanced approaches, that demonstrate the effectiveness of the proposed methodology.