Performance Analysis of Deep Learning Models for Re-identification of a Person in a Public Surveillance System

Abstract

Nowadays, a huge network of cameras is deployed in public locations, generating enormous amounts of video data. These data are monitored manually and can only be accessed if the necessity arises to verify the facts. Automating the system can increase the quality of monitoring and can be beneficial for high-level surveillance activities such as person recognition, suspicious activity detection, and unwanted event prediction for timely notifications. Person re-identification (ReID) tries to identify and track the activities of persons across multiple cameras. The common challenges faced are non-overlapping cameras, occlusion, large changes of viewpoint, shadows, various scales, cluttered background, pose, and illumination across different fields of view. This chapter proposes an automatic real-world video surveillance system model that can track and re-identify multiple persons from a single-camera tracking environment. The proposed system compares the effect of different benchmarked deep neural networks on MARS and iLIDS-VID video re-identification dataset. DenseNet121, SEResNeXt50, Resnet50M, Resnet-50, and inceptionresnetV2 are the well-known neural networks used in the various image or video analytics domains. We train these models to analyze the effect in the re-identification domain. Further, we also analyzed the model based on the effect of hyper-parameters like learning rate, number of epochs, dropout rate, loss function, and pooling strategies. The results obtained by the training DenseNet121 model using the MARS dataset are 66.0% and the 24iLIDS-VID Dataset is 76.2%. The system’s benefit is that it does not require a pre-stored database of individuals in advance at the time of testing for person recognition; it will create the dataset in real time. It will be also beneficial for crime control and prevention. Based on the experimental results produced, we conclude that standard deep learning techniques did not perform well due to various challenges including computational requirements, adequate training, and environmental issues. Hence, there is a need to investigate a suitable solution for a real-world automatic surveillance system.