Research On Lightweight Sensor Traffic Pattern Recognition Algorithm Based On Deep Learning

With socio-economic development and a growing urban population,urban traffic pressure is increasing and traffic safety and congestion problems are becoming increasingly prominent.Therefore,traffic pattern recognition,as one of the important branches of human behaviour recognition,has become an integral part of modern urban transport systems.However,most current traffic pattern recognition systems use high power consumption sensors(e.g.GPS and Wi-Fi sensors)which are unable to collect data in the absence of a signal and consume significant amounts of power.In addition,most traffic recognition models use machine learning algorithms,which require manual feature extraction and suffer from problems such as unstable recognition,complexity,time consumption and low recognition accuracy.Although some researchers have proposed deep learning-based traffic pattern recognition algorithms,these algorithms still suffer from high power consumption,low accuracy and high training time complexity.Therefore,in this thesis,algorithms based on a time-domain convolutional attention mechanism and a recurrent neural network based on an attention mechanism are constructed to investigate the recognition of eight traffic patterns(stationary,walking,running,bicycle,car,underground,bus and train)of users.Firstly,this thesis examines the background and significance of traffic pattern recognition.Research in traffic pattern recognition can alleviate the ever-increasing problem of traffic congestion and traffic accidents.This thesis summarises the current state of the art in traffic pattern recognition research data and the current state of the art in research methods.Specifically,research data for traffic pattern recognition can be divided into data based on wireless networks and data based on lightweight sensors in mobile phones.The research data can be divided into traffic pattern recognition based on lightweight sensor data from wireless networks and mobile phones;and the research methods can be divided into traffic pattern recognition based on machine learning algorithms and traffic pattern recognition based on deep learning algorithms.Secondly,a traffic pattern recognition algorithm(TFPA)based on a time-domain convolutional attention mechanism is constructed in this thesis.The algorithm uses lightweight sensors in smartphones(including gyroscopes,linear acceleration sensors,geomagnetic,gravity,acceleration and barometric sensors)to extract deep elemental features using a multi-channel approach using time-domain convolutional networks,and processes the data in a self-fusion manner before further correcting the data information using a multi-headed attention mechanism.In addition,network degradation is addressed by adding residual connections.The algorithm was also evaluated using the Sussex-Huawei Campaign(SHL)traffic pattern recognition dataset.The experimental results showed that the model recognized traffic patterns with an accuracy of 95.36%,which was significantly better than the nine baseline models.In the recognition of each traffic pattern,the algorithm in this thesis has an accuracy and recall rate higher than 91%,verifying its accuracy and robustness.To address the problem of high training time complexity,this thesis also constructs a time-domain convolutional attention mechanism traffic pattern recognition algorithm based on the bagging method,which can effectively reduce the training time.Finally,with the continuous research on traffic pattern recognition algorithms,this thesis constructs a more lightweight attention mechanism based recurrent neural network model(CPAG)for the problem of high training time complexity of TFPA.The model uses the idea of chunking,aiming to extract deep elemental features in the data,and consists of a local feature extraction module and a global feature extraction module,using different channels to learn the information at the same time.The local feature extraction module consists of a convolutional neural network,an activation layer and a pooling layer,while the global feature extraction module consists of an attention mechanism and a fully connected layer,incorporating positional coding to establish sequential relationships between elements.After feature extraction,the model uses a self-fusion approach and incorporates gated recursive units for recognition.The algorithm reduces the complexity of high-dimensional inputs and uses two types of sensor data(gyroscope sensors,linear acceleration sensors)to achieve a small number of sensors to identify traffic patterns.Experimental results show that the accuracy of the model is higher than that of the compared baseline model in both cases,and achieves the goal of reducing training time complexity.The experimental results show that the accuracy of the model is all higher than that of the compared baseline model.The aim of this thesis is to explore the optimisation of traffic recognition algorithms in order to improve traffic recognition accuracy and applicability.Through the experimental study of different traffic recognition algorithms,this thesis concludes that a high recognition accuracy can be obtained by using a traffic pattern recognition algorithm based on the time-domain convolutional attention mechanism.If resources are sufficient,a time-domain convolutional attention mechanism traffic pattern recognition algorithm based on the bagging method may be considered.If the algorithm needs to be embedded in a smart terminal in the future,the use of an attention-based mechanism circular convolution traffic pattern recognition algorithm could be considered.