Liu Jingna, Han Xiang, Wang Yuning, Wang Xiaohe
Hebei Tonghua Highway Materials Co., Ltd., Sichuan Yunzhi Technology Co., Ltd., School of Management and Economics, Beijing Institute of Technology
Abstract: Combined with satellite navigation and deep learning, this paper proposes to use a deep neural network model to mine features from the trajectory data of trucks in transit to distinguish the empty driving and full load states of trucks. The accuracy of the model in judging empty and full load on the verification set can reach more than 85%, which can be extended to further determine whether the truck is overloaded, and provide effective support for the transportation management department to accurately control the overload.
Keywords: satellite navigation; deep learning; Truck trajectory;
About author:LIU Jingna (1986—), female, from Shijiazhuang, Hebei Province, engineer, research direction is electronic engineering.
0 Introduction
With the large-scale application of the mainland Beidou satellite navigation system in the field of transportation, the corresponding transportation management platform has a sufficient amount of data accumulation after years of operation, and shows a steady growth trend, its effective use of truck operation data in transit, improve the efficiency of transportation management and scheduling, reduce the labor intensity of management personnel, replace the traditional manual mode with artificial intelligence algorithms, belongs to the Beidou satellite positioning system, artificial intelligence and bulk cargo transportation combined with cross-field research. Based on the data collected in actual business scenarios, this paper is modeled by artificial intelligence deep neural network, and finally obtains a recognition standard that can effectively determine the freight status of driving trucks, with a recognition accuracy of more than 85%.
In the process of bulk cargo transportation, the shipper has higher requirements for the safety of the goods in transit during transportation. At present, the Ministry of Transport is fully implementing and adhering to the development of "Internet +" efficient logistics, and promoting the solution of deep-seated contradictions and problems accumulated in the road freight industry over the years through the development of network freight. The network freight model not only improves the efficiency of road transport, reduces transportation costs, but also reduces carbon emissions. But it also brings some social problems, such as how to ensure the safety of transported goods, prevent smuggling, cross-trade and other issues. In this paper, 1 000 heavy truck trajectories and a total of more than 300,000 trajectory points were randomly selected as training and test data using the transportation management system developed by Hebei Tonghua Highway Materials Co., Ltd., and the artificial intelligence deep neural network was used for modeling. After the intelligent model is solidified, it is deployed as an intelligent algorithm to the truck management information system, and combined with the driving track and time stamp of the truck, it is used to determine whether the goods are loaded and unloaded at the designated place during transportation, and whether the behavior of unloading in the dangerous area (cross-shipment) has occurred. When the above behaviors are detected, the model can raise alarms in time to remind managers and consignees to check the corresponding trucks.
1 System Build
The components related to this paper in the whole vehicle management system include: vehicle Beidou positioning terminal, server data receiving end, server data storage end.
The vehicle-mounted Beidou positioning terminal is installed in the glove box of the co-pilot of each truck, and the equipment is continuously powered by power. When it is in a stationary state, it sends a current location data to the server every minute, including longitude, latitude, altitude, speed, driving direction, and generated timestamp. In the state of motion, a current position data is sent to the server every 2s.
The server data receiver runs on the cloud server, which is responsible for receiving and parsing the data sent by the vehicle terminal, and sends the parsed data to the data storage terminal through the message queue.
The server data storage end is responsible for storing the messages received by the server data receiver to the hard disk according to the rules of vehicle and timing.
2 State of the art
Truck status recognition technology refers to the technology that determines the state of automobile freight (such as empty/full load) through continuous truck trajectory data, and is a trajectory classification technology. Trajectory classification is used in many fields, including smart city traffic planning, recognition of user trajectory traffic patterns (such as identifying vehicle categories, identifying suspicious vehicles, identifying the operating status of taxis, etc.), as well as classifying and judging trajectories, identifying abnormal trajectories, etc. [1], which are closely related to daily life. Krumm et al. [2] proposed a spatiotemporal trajectory classification algorithm based on Wi-Fi data, using the hidden Markov model to divide the trajectories into two categories: stay and motion. Yin et al. [3] proposed a DBN model, in which the lowest layer is the Wi-Fi signal, and based on this layer, the user activity category is finally obtained by mining and analyzing the upper layer [4].
With the wide application of the BeiDou satellite positioning system, a large number of studies based on motion trajectory classification have begun to be applied to massive positioning data, and researchers have established classification models through statistical analysis, feature extraction, and other methods, and obtained high accuracy [5]. Gonzalez et al. [6] used GPS data and neural network principles to automatically classify and judge traffic modes, and their important contribution was to propose a key point algorithm to filter part of the GPS data to refine the core information. Secondly, the researchers extracted data features and key points such as average velocity and dwell time as the input of the neural network, and obtained recognition results with high accuracy [7]. Nowadays, smartphones, tablets, wearable devices, new energy vehicles, trucks, and online car-hailing are all continuously generating a large amount of location data, which greatly increases the opportunity for researchers to obtain massive location data in various scenarios, which is conducive to the evolution and verification of trajectory research algorithms, and promotes the improvement of prediction accuracy in related fields. As⁃semi et al. [8] collected data using the mobile app ATLAS II (Advanced Travel Log⁃ging Application for Smartphones II, ATLAS II), which runs in the background and continuously collects information such as latitude and longitude, speed, direction, and timestamp of the device, and stops recording data when the user's dwell time exceeds the set dwell threshold. In addition to analyzing the data collected by the sensor, Wang et al. [9] found that Call Detail Records (CDR) can also provide a large amount of user location information, which is generated by users' behaviors such as sending and receiving text messages, making phone calls, and mobile Internet access, which are easy to obtain. Through the above data, researchers can analyze the relevant information of the user, such as user identification, origin, destination, start time, end time, etc.
However, the classification and prediction of spatiotemporal trajectories is a comprehensive and complex problem. At present, the difficulties faced by this field can be divided into the following categories: First, the shortcomings of the data itself lead to the limitation of the study of spatiotemporal trajectories. For example, the error or even wrong data of the original trajectory data itself (the accuracy of Beidou positioning is reduced due to weather and geographical location, the wrong trajectory point caused by the multipath effect caused by the reflection of the urban glass curtain wall, and the position drift caused by the signal interference of terminal equipment, etc.); Although the overall sample size of trajectory data is very large, the amount of data refined to specific scenarios, specific regions and specific participating groups is relatively small, so more accurate feature item selection and longer-term follow-up research are required, and a large number of in-depth vertical domain expertise is also needed. Due to the differences in the habits and individuals of different users, for example, the characteristics of starting and braking will be different due to the different driving habits of different drivers, and the trajectory characteristics of the same scene are also diverse, so it is also a difficult and challenging problem to study the trajectory classification problem with the characteristics of individual users [10].
3 Model Building
In this project, the focus of identification is on capturing the motion characteristics of trucks based on the motion characteristics of trucks. Since the characteristics of truck motion are very similar in some situations (e.g., starting, braking, turning, high-speed overtaking, etc.), the key problem in this project is how to extract the most discriminating features of the freight trajectory. Deep learning models are capable of feature learning, abstracting features layer by layer and extracting key features, and being able to distinguish the nuances of model inputs. Among them, convolutional neural network [11] is a special form of artificial neural network, which is one of the core algorithms in the field of deep learning [12], and is also used in image semantic segmentation, scene classification, image saliency detection and other problems. The project team used a multi-layer convolutional neural network to build a model and automatically learn the characteristics of freight trajectories.
The structure of the model is shown in Figure 1, which consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer is the continuous motion trajectory point data containing four channels such as longitude, latitude, velocity, and angle, and the network can be compatible with multiple convolution and pooling layers. After the multi-channel features are extracted by convolution and pooling, they are reshaped into one-dimensional long vectors, which are mapped into two neurons through the fully connected layer, which respectively represent the predicted probability of the empty full load state, and achieve the purpose of empty full load recognition through their sizes. The following is a detailed description of the data processing and the structure of each layer of this model.
3.1 Trajectory data preprocessing
3.1.1 Trajectory threshold processing
If outliers are not screened out in time in the collection of trajectory data, the analysis efficiency and accuracy of effective trajectory data mining will be seriously reduced. Therefore, in the preprocessing stage, it is necessary to first check and screen out the abnormal data, including suspicious movements and record abnormal trajectories, and then eliminate the error data generated by the superposition of multiple error sources. Examples include anomalous random errors due to satellites, atmospheric interference, and multipath signal reflections. In the project, the trajectory data is detected and removed by performing the following steps to clean and remove inaccurate or abnormal trajectory data points: first, identify and remove those trajectory data points with timestamps much larger than the next time point; Secondly, according to the range of moving speed, the trajectory points with abnormally high velocity that do not conform to the movement speed are removed. Finally, considering the maximum angle change, the track points with high angle changes outside the value range are removed, and the longitude and latitude track points that exceed the threshold range are identified and removed.
Figure 1 Neural network architecture and input and output Download the original image
3.1.2 Trajectory smoothing
In the previous data preprocessing, the project team has thresholded the trajectories to eliminate random errors. At this stage, the project team smooths the remaining trajectory data after screening. In this model, the Savitzky-Golay filter is applied to all trajectory segments, and the appropriate filtering parameters are determined according to the average trend of the sequence curve, and the polynomial is used to realize the filtering method of least squares fitting in the sliding window.
3.2 Convolutional neural network structure
(1) Convolutional layer. The convolutional layer is the core structure of the model, which is composed of multiple convolution units and a set of filters. Each neuron output from the convolutional layer is fully connected to the receptive field of the previous layer, and its size is equal to the size of the filter. The output value of the neuron is derived from the dot product of the filter's parameter values corresponding to the receptive field. The weight parameters of each convolutional kernel are optimized and solved by the backpropagation algorithm. Drill down the data in this layer to get more abstract eigenvalues. The convolution operation uses a nonlinear activation function to map and normalize the output to a set range of values [13]. When the feature intensity of a region does not reach the set threshold, the output of the activation function is 0, indicating that the extraction method of the corresponding feature item can not effectively activate the feature of the region, or indicating that the feature of the region is very weak, and only a small number of neurons need to change the state without global adjustment, which reduces the degree of information coupling, and then reduces the difficulty of network training, although the output feature is sparsely expressed, However, the activated output can still maintain the original expressive ability, that is, the area that is not related to the feature, and will not affect the training of the feature extraction method. At the same time, in order to avoid the effect of the ReLU function as a non-zero mean function affecting the fast convergence effect of the network, the output is normalized in this model.
(2) Pooling layer. In this stage, the feature dimensionality reduction process is carried out to compress the number of data and parameters, reduce overfitting and other processing, and improve the fault tolerance rate of the model. The eigenvalues are divided into regions with the regions where the corresponding matrices are located, and new eigenvalues are calculated for each segment. In this model, the maximum pooling is used to divide the input feature map into several region blocks of equal size, and the maximum value of the elements in each region is taken. On the one hand, this sampling method can suppress the noise and reduce the deviation of the estimated mean caused by the error of convolutional layer parameters. On the other hand, it can improve the significance of the feature map in the region (the filtered maximum) as a new feature value output.
(3) Fully connected layer. This layer acts as a classifier, and the features of the previous layer are weighted and summed, and the feature space is mapped to the sample label space through linear transformation. This is in contrast to convolutional layers, which map data inputs to hidden layer feature spaces.
(4) Output layer. The output layer uses the Softmax activation function for classification prediction, which is used to output the probability value that a certain sample belongs to different freight states. The output layer and the previous fully connected layer form a shallow Soft⁃max classifier, which is generalized on the basis of logistic regression. The output layer contains two values, which represent the probability estimation of the no-load and full-load states, respectively, and finally judge the no-load state according to the size of the two.
4 Test Results
The project team used TensorFlow to complete the programmatic construction of the above model, and imported 1 000 randomly selected trajectories and more than 300,000 trajectory points into the model as training data and validation data for training and validation. 80% of the dataset is used as the training set and 20% of the dataset is used as the validation set. After 1 000 rounds of iterative optimization, the model finally obtained: the accuracy rate of discriminating empty full load on the training set is 89.75%, and the accuracy rate of discriminating empty full load on the validation set is 92.96%. The accuracy on the validation set is quite close to that on the training set, indicating that the AI algorithm does not have the problem of overfitting and can be put into actual scenarios for automatic discrimination. The results of training and validating the model with TensorFlow are shown in Figure 2.
5. Application scenario expansion
In this application scenario, the data used for training is the Beidou driving trajectory data of the asphalt truck, which is mainly in two states: empty (0t cargo) and full load (not overloaded, about 33t cargo). In order to meet the requirements of the traffic supervision department for accurate overloading, it is only necessary to further import the Beidou driving trajectory data of the overloaded trucks inspected by the traffic management department into the model, capture the motion characteristics of the overloaded trucks through the method of neural network deep learning, and solidify the features into the algorithm. After the algorithm is deployed online, it is only necessary to continuously pass the Beidou driving trajectory data of the truck into the artificial intelligence algorithm, and the automatic algorithm can be used to timely and accurately judge the status of the truck being empty, fully loaded and overloaded, so as to efficiently assist enterprises and cargo owners to supervise the safety of goods in transit, and help the traffic management department to supervise whether the truck is overloaded.
Figure 2 Model training and validation results Download the original image
6 Conclusion
This project uses the Beidou satellite positioning system and artificial intelligence technology to automatically identify the empty, full and overloaded states of trucks, with a success rate of more than 85%. This technology is of great significance to the automatic scheduling of trucks in the freight industry and the precise management of the traffic management department. At the same time, such applications will also accelerate the digital transformation of the logistics industry, promote the rise of new business formats such as smart logistics, platform economy, and fully automated transportation, and make more and more logistics links usher in subversive changes in disorganization and reconstruction.
bibliography
[1] JG Lee, J Han, X Li. Trajectory Outlier Detection:A Partition-and-Detect Framework[C]//2008 IEEE 24th International Conference on Data Engineering. IEEE,2008:140-149.
[2] J Krumm, E Horvitz. LOCADIO:Inferring Motion and Location from Wi-Fi Signal Strengths[C]//Mobile and Ubiquitous Systems:Networking and Services, 2004.MOBIQUITOUS 2004. The First Annual International Conference on. 2004:4-13.
[3] J Yin, X Chai, Y Qang. High-Level Goal Recognition in a Wireless LAN[C]//Conference on Nineteenth National Conference on Artificial Intelligence. 2004:578-584.
[4] Zhao Zhujun, Ji Genlin. Journal of Geo-information Science,2017,19(3):289-297.)
[5] Y Zheng. Trajectory Data Mining:An Overview[J]. ACM Transactions on Intelligent Systems and Technology(TIST), 2015, 6(3):1-41.
[6] Paola A Gonzalez, Jeremy S Weinstein, Sean J Barbeau,et al. Automating Mode Detection Using Neural Networks and Assisted GPS Data Collected Using 1663GPS-enabled Mobile Phones[C]//15th World Congress on Intelligent Transportation Systems:16-20.
[7] Xiao Yanli, Zhang Zhenyu, Yang Wenzhong. Chinese Journal of Intelligent Systems,2014(5):536-543.)
[8] B Assemi, H Safi, M Mesbah, et al. Developing and Validating a Statistical Model for Travel Mode Identification on Smartphones[J]. IEEE Transactions on Intelligent Transportation Systems, 2016, 17(7):1920-1931.
[9] H Wang, F Calabrese, GD Lorenzo, et al. Transportation Mode Inference From Anonymized and Aggregated Mobile Phone Call Detail Records[C]//13th International IEEE Conference on Intelligent Transportation Systems. IEEE, 2010:318-323.
[10] S Xing, H Caceres, H Tong, et al. Online Travel Mode Identification Using Smartphones With Battery Saving Considerations[J]. IEEE Transactions on Intelligent Transportation Systems, 2016, 17(10):2921-2934.
[11] Kunihiko Fukushima, Improvement of Pattern Separation Ability of Cognitron [J].Trans. IEICE, 1979, 62(10):650-657.
[12] Fang Hao. Research on traffic mode recognition based on convolutional neural network[D].Beijing:Beijing Jiaotong University,2019.
[13] Schmidhuber Jürgen. Deep Learning in Neural Networks:An Overview[J]. Neural Networks:The Official Journal of the International Neural Network Society,2015(61):85-117.
Disclaimer: We respect originality and also focus on sharing. Some of the content comes from the Internet, the copyright belongs to the original author, only for learning and reference, it is forbidden to use it for commercial purposes, if you inadvertently infringe on the intellectual property rights of any media, company, enterprise or individual, please contact to delete (email: [email protected]), and the content of this headline push only represents the author's point of view, and has nothing to do with the headline operator, the authenticity of the content please readers identify by themselves, this headline does not assume any responsibility.