Research On Vehicle Load Pattern Recognition Algorithm Based On Deep Learning

Road freight is an important part of China’s transport industry,which is of great significance to economic development and people’s livelihood protection,etc.However,the phenomenon of string exchange of goods,high empty rate of long-haul transport backhaul,and high risk of traffic accidents on the way of freight transport restrict its further development,and the above problems can be alleviated to a certain extent by real-time detection of truck load status.In recent years,China’s truck weighing technology has made good progress,but there are still problems such as large footprint,high promotion cost and low popularity.And due to the popularity of vehicle-mounted positioning equipment,it makes the truck GPS track data inexpensive to obtain,which provides the data basis for identifying the truck load status.On the other hand,with the development of deep learning,it has been applied to traffic pattern recognition,ship track classification,vehicle type recognition and other spatio-temporal trajectory classification problems.Thus,this thesis intends to use GPS trajectory data of trucks to study deep learning-based truck load state recognition algorithm,and the main research contents are as follows.（1）In order to solve the problem that the traditional method relies on feature engineering and the model accuracy is greatly affected by the degree of expert experience,a convolutional neural network-based truck load classification method is proposed.The method first preprocesses the truck trajectory data and generates the feature matrix through the steps of kinematic feature extraction,threshold processing and trajectory segmentation.Subsequently,convolutional neural network is used to further extract high level features and discriminate the truck load status by Softmax classifier.The experimental results show that the method can effectively identify the truck loading status with an accuracy rate of 81.2%.（2）In view of the existence of certain correlation between different trajectory features,in order to better model their correlation,a truck load classification method based on temporal imagery is proposed on the basis of the above method.The method applies the Gram’s angle fielding method to transform the trajectory features into images to further improve the representation of the trajectory features;and introduces two attention mechanisms to achieve the learning of correlation between different kinematic features by obtaining attention weights.The comparison experiments show that the method can effectively improve the accuracy of truck load status recognition.（3）In order to solve the problem of the large amount of trajectory data to be processed in the process of practical application,a truck load status detection system is constructed.The system adopts stream data processing architecture,realizes the processing of large amount of trajectory data based on Flink technology,and realizes the batch deployment of prediction models by using container technology.In addition,the system has the functions of real-time processing of trajectory data,truck load status identification,model training and deployment,and trajectory data query.The test results show that the system can meet the needs of large-volume truck load status detection.In this thesis,two truck load recognition algorithms are proposed and a truck load status detection system is built based on them.Experiments were carried out on the collected truck trajectory data,and the experimental results show that the algorithms can identify the truck load status better.And at the end,tests were carried out using the trajectory data to verify the system functions.