Towards the development of automated traffic monitoring system

Abstract

Traffic management refers to the process of controlling and regulating the flow of vehicles, pedestrians and other road users to ensure safety, efficiency and convenience on roads. Traffic rules are essential to regulate traffic movement and ensure road safety. In the past few years, road accidents have increased significantly within metro cities and on highways. Automated traffic monitoring system may be developed to monitor traffic rule violation cases and help the administration to ensure road safety. For motorcycle riders, wearing a helmet is one of the most effective ways to protect themselves and reduce the severity of injuries in the event of an accident. It has been observed that the clarity of images and videos captured by surveillance camera are often reduced due to bad weather conditions like fog, mist, rain or pollution. Hence, restoration of hazy images and videos can improve the performance of automated traffic monitoring system. In this thesis, an end-to-end morphological network fusion attention based image dehazing method is proposed. Further, image deraining technique is introduced by applying Convolutional Neural Network (CNN) with attention mechanism named Style-based Recalibration Module (SRM). Additionally, Surveillance video dehazing method is proposed by using a lightweight U-Net architecture. In each layer of U-Net, adaptive frequency based channel attention module is applied that enhances the model accuracy significantly. This thesis also proposes an automated helmet detection framework for single and multiple motorcycle riders using CNN. To assist automated traffic monitoring system, video captions are generated corresponding the traffic video scene by proposing a deep neural network-based model. In the proposed encoder-decoder based model, CNN along with optical flow guided approach are used for visual feature extraction in the encoder part. To design decoder, Recurrent Neural Network (RNN) based Long-Short-Term-Memory (LSTM) with Soft Attention technique is applied to enhance the mode