The movement of hazardous goods involves risks from leaving the warehouse to being transported by vehicle to the destination. Monitoring accidents, vehicle movement dynamics, and how often vehicles pass through a particular area are crucial in supervising vehicles' transportation process of hazardous goods.
The GPS placement of the hazardous goods vehicle serves as the primary foundation for dynamic vehicle movement monitoring. The number of times and continuous detection time of the vehicle identified by the camera may be used to determine the specific frequency of passing through a particular site and accident conditions of the hazardous goods vehicles. The accuracy of hazardous goods vehicle identification is significantly impacted by ambient characteristics such as illumination, messy background, and partial occlusion.
Image-Based and Deep Learning-Based Vehicle Identification Techniques
The two basic vehicle identification techniques are image-based and deep learning-based.
The image-based detection approach primarily identifies vehicle targets through vehicle image characteristics and directional gradient histogram features. The fundamental drawback of vehicle detection systems based on vehicle image texture and edge features is that lighting and vehicle integrity significantly impact them. However, as deep learning continues to advance, an increasing number of researchers are looking at the topic of vehicle detection using deep learning.
In this study, vehicle detection was implemented using the deep learning technique. To achieve quick and precise vehicle recognition, researchers enhanced the training phase of the deep learning EfficientDet model and constructed a phased training model.
First, the hazardous goods vehicles are trained using an efficient deep learning model, and then the trained model is utilized to identify the hazardous goods vehicles. The deep learning-based target identification network mainly includes CenterNet, SpineNet and cascade R-CNN.
Different Deep Learning Network Models
The EfficientDet deep learning network model has the best detection performance, followed by the cascade R-CNN_ResNet deep learning network model. The EfficientDet-D7x network has the most significant detection performance of the EfficientDet series, with the highest recognition accuracy of 55.1, followed by EfficientDet-D3. However, due to its excessive complexity, the EfficientDet-D7x network cannot be utilized to meet the need for real-time vehicle detection. Therefore, EfficientDet-D3 was used in this study as the vehicle detection network.
Training of the Model
Data sets of vehicle images were gathered in various settings and at various times. The training, verification, and test data sets were created from this data set, i.e. 146 test data sets, 211 verification data sets, and 2387 training data sets. Before training, the configuration file's model-specific settings, such as related data reading path, initial learning rate, batch size, and the number of classes, are modified according to the data set's properties.
Case Study for Research Validation
Researchers used 146 test data sets to identify hazardous goods vehicles and compare the detection outcomes with CenterNet, cascade R-CNN, and EfficientDet-D7x techniques to validate the effectiveness of the proposed approach. The Hanyang district of Wuhan has four warehouses for hazardous goods. Using trained deep learning models, researchers identified hazardous goods vehicles on cameras in 10 sites in this region.
Significant Findings of the Study
In this study, a hazardous goods vehicle identification model based on the Efficientdet-d3 model was constructed, and a hazardous goods vehicle detection approach based on deep learning was presented. The setup of phased training and learning parameters was supplied according to the change in total loss value during the training stage of the Efficientdet-D3 model to increase the training efficiency of the detection model.
Comparing Cascade R-CNN, CenterNet and Efficientdet-D3 Models
The detection model employs the fewest parameters and has the lowest computational cost compared to approaches based on cascade R-CNN and CenterNet.
This technique entirely beats the cascade R-CNN method in the detection time for hazardous goods vehicles, which is comparable to that of the CenterNet method. The detection accuracy of these three approaches is more or less the same. Detection accuracy, time consumption, and computing complexity indicate that the approach presented in this study is superior to the other two.
Results of Case Study
Hazardous goods vehicles in various scenarios are evaluated using this detection model. The findings demonstrate that this system can reliably identify hazardous goods vehicles in various settings.
The Wuhan Petrogoods Company used the deep learning model built in this study to identify hazardous goods vehicles in four different divisions. The results of the experiments demonstrate that this method's accuracy is more than 90%.
This study examines the detection of hazardous goods vehicles in the areas around four hazardous product facilities in Hanyang district of Wuhan. It calculates the weekly number of hazardous goods vehicles that travel through each area.
The danger level of each segment around the hazardous goods warehouse can then be easily observed on the map by calculating the risk level for each sector based on the number of times hazardous goods vehicles travel through each region.
Reference
Qing An, Shisong Wu, Ruizhe Shi, Haojun Wang, Jun Yu and Zhifeng Li (2022) Intelligent Detection of Hazardous Goods Vehicles and Determination of Risk Grade Based on Deep Learning. Sensors. https://www.mdpi.com/1424-8220/22/19/7123/htm
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.