Analysis of Deep Learning Models for Early Action Prediction Using LSTM

Abstract

AbstractVideo surveillance is being increasingly adopted for ensuring safety and security both in public and private places. Automated prediction of abnormal events like theft, robbery, murder etc., from continuous observation of surveillance videos is a multidisciplinary study involving computer vision, deep learning and artificial intelligence. Deep learning-based video analysis and categorization is a most researched topic. Many deep learning models based on long short-term memory (LSTM) are proposed for automated prediction of abnormal events. This work does a comparative analysis of six LSTM-based deep learning models for abnormal event prediction from surveillance videos. Deep learning models of ResNet, VGG16, VGG19, 3DCNN, Inception V2 and Inception-ResNet-V2 are combined with LSTM for prediction of abnormal event from past observation of events in the video stream. These six models are run against different benchmarked abnormal event detection datasets and performance is compared in terms of accuracy, loss and execution time. It is observed that Inception-ResNet with LSTM provides training accuracy of 80% and test accuracy of 80% compared to other models.Keywords3DCNNInceptionInception-ResNetVGG16VGG19ResNet