Uni-Dimensional Autoencoder Reinforced Multilayer Perceptron Network for Individual Behavior Detection

Abstract

In recent years, due to the increasing number of public security incidents, the field of individual behavior detection has made great progress. Among them, MLP method is representative, but its defects are also very obvious. this paper proposes an autoencoder fusion MLP based novel network structure for behavior detection, which can significantly improve the recognition accuracy. The proposed network extracts the color-based features from the video, outputs and compressed the features as a one-dimensional vector with autoencoder, and finally input the parameters into the fully connected layer for the classification of abnormal behaviors. The proposed network achieved the accuracy of 67% on the UCF-Crime data set, and significantly enhanced the accuracy on simple data sets. The experimental results indicate the autoencoder achieves promising performance on individual behavior recognition and potentially on the crowd behaviors.