Detecting Abnormal Event in Traffic Scenes using Unsupervised Deep Learning Approach
K. Sindhu Meena, A. Viji, J. Joshan Athanesious, V. Vaidehi · 2019
In today's digital age, detecting abnormal events in the traffic scene is a more significant problem in video surveillance. The proposed scheme is adopted for detecting abnormalities such as wrong side driving, people crossing the road illegally, vehicle moving in pedestrain pathway and which uses spatio-temporal attributes. Although numerous machine learning approaches are described in existing literature focusing on pixel-wise differnce with less performance in identifying the abnormalities in video traffic scenes. In this paper, a novel frame work for identifying abnormal detection using unsupervised deep learning algorithm. The scheme uses joint based methodology (ConvLSTM with kmeans) with a combination of reconstruction and clustering loss respectively. The convolutional neural network with LSTM is used to detect and recognize anomalies in traffic scenes with refernce of previous frame information. To identify the normal or abnormal events, the proposed work is composed of training and testing phase. The proposed model is implemented in real time and it is found that the accuracy rate for detecting abnormalities 93.02% abnormality than state-of-art technique.