KR-102962575-B1 - DEEP LEARNING-BASED GPS PATH PREDICTION METHOD AND SYSTEM FOR MOVING OBJECTS

Abstract

The present invention is characterized in that it comprises: a GPS path receiving unit that receives GPS path data, which is a set of GPS coordinates collected at regular intervals; a learning unit that trains a deep learning model by taking a coordinate at an arbitrary time interval from training data composed of GPS path data as an input value, and taking a coordinate immediately after the arbitrary time interval or a coordinate after a certain time interval from the arbitrary time interval as an output value; and a path prediction unit that predicts a coordinate immediately after the last coordinate of the GPS path data received by the GPS path receiving unit or a coordinate after a certain time interval from the last coordinate using the trained deep learning model.

Inventors

• 이규철
• 윤승원

Assignees

• 충남대학교산학협력단

Dates

Publication Date

20260508

Application Date

20220427

Priority Date

20220425

Claims (7)

1. In a GPS path prediction system using deep learning, A GPS route receiver that receives GPS route data, which is a set of GPS coordinates collected at regular intervals; A normalization unit that scales the above GPS path data to a value between 0 and 1; A learning unit that takes coordinates at an arbitrary time period as input values from training data composed solely of the above GPS route data, calculates the Root Mean Square Error (MSE) by taking coordinates immediately after the said arbitrary time period or coordinates after a certain time period from the said arbitrary time period as output values, and trains a Long Short-Term Memory (LSTM) deep learning model using a supervised learning method by applying an Adam optimizer; A path prediction unit that predicts the coordinate immediately after the last coordinate of the GPS path data received by the GPS path receiver, or the coordinate after a certain time from the last coordinate, using a trained LSTM deep learning model; and It includes a denormalization unit that rescales the predicted coordinates calculated by the path prediction unit to the range prior to scaling by the normalization unit; A GPS path prediction system characterized by the above-described learning unit calculating the root mean square error (MSE) using the difference between the output value and the predicted coordinate value, and optimizing the deep learning model by applying an Adam optimizer.
2. In Article 1, The above learning unit is, A GPS path prediction system characterized by using Long Short-term models (LSTM) as a deep learning model.
3. delete
4. delete
5. delete
6. A GPS route receiving step for receiving GPS route data, which is a set of GPS coordinates collected at regular intervals; A normalization step for scaling GPS route data received in the above GPS route reception step into a value between 0 and 1; A learning step of using coordinates at an arbitrary time period as input values in training data composed solely of the above GPS route data, calculating the Root Mean Square Error (MSE) by using coordinates immediately after the said arbitrary time period or coordinates after a certain time period from the said arbitrary time period as output values, and training a Long Short-Term Memory (LSTM) deep learning model using a supervised learning method by applying an Adam optimizer; A path prediction step that predicts the coordinate immediately after the last coordinate of the GPS path data received in the GPS path reception step, or the coordinate after a certain time from the last coordinate, using a trained LSTM deep learning model; and It includes a denormalization step for rescaling the predicted coordinates calculated in the path prediction step to the range prior to scaling in the normalization step; A GPS path prediction method characterized by the above learning step calculating the root mean square error (MSE) using the difference between the output value and the predicted coordinate value, and optimizing the above LSTM deep learning model by applying an Adam optimizer.
7. delete

Description

Deep Learning-Based GPS Path Prediction Method and System for Moving Objects The present invention relates to a deep learning-based GPS path prediction system for a moving object, and in particular, to a deep learning-based GPS path prediction system for a moving object in which a deep learning model learns the GPS path of a moving object, such as a car and a pedestrian, and can predict the next path. Moving objects such as automobiles and pedestrians move and have a trajectory. These trajectories have patterns, and identifying such patterns to predict a pedestrian's next path is effective for research such as risk prediction. By predicting a pedestrian's path, various accidents can be prevented by providing warnings to objects likely to be close to dangerous situations or dangerous individuals. In the past, mathematical probability-based prediction models were primarily used for path prediction, such as AR, MA, ARMA, and ARIMA models, which provide regularity to irregular time-series data through a form of normalization, and Hidden Markov Models (HMM), which compensate for the shortcomings of Markov chains by hiding the flow of states. However, these models have the problem that it is difficult to achieve high-performance predictions because they cannot learn a large amount of paths. Unlike existing probability-based prediction models, path prediction research using deep learning models can learn a large amount of data, and such deep learning-based prediction systems can provide different prediction data depending on the input data (GPS, camera). Research that detects objects and predicts their paths using cameras, CCTVs, etc., has the limitation that path prediction is only possible within the camera's range, so it is not suitable for research on predicting the paths of moving objects traveling over a wide area. Path prediction research using GPS sensors allows for the learning of moving object paths over a wider range because there are no limitations on the object's radius of movement. Such research utilizing GPS sensors can be divided into studies that preprocess data by mapping it onto a grid map for training, and studies that train the GPS data sequence itself. While research using grid maps has the advantage of facilitating model training due to reduced complexity of GPS data, path prediction becomes impossible if the moving object moves beyond the boundaries of the gridded map. Therefore, to solve these problems, there is a need for a deep learning-based GPS path prediction system for a moving vehicle that can predict the next path regardless of the path range. This project (result) is the result of the Local Government-University Cooperation-based Regional Innovation Project conducted in 2021 with funding from the Ministry of Education and support from the National Research Foundation of Korea (2021RIS-004). This results was supported by "Regional Innovation Strategy (RIS)" through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(MOE)(2021RIS-004) Figure 1 shows a configuration diagram of a GPS path prediction system according to an embodiment of the present invention. Figure 2 shows the configuration of training data and test data for predicting the coordinates immediately following GPS path data of a learning unit according to an embodiment of the present invention. Figure 3 shows the configuration of training data and test data for predicting coordinates after a certain period of time of GPS path data of a learning unit according to an embodiment of the present invention. Figure 4 shows the structure of an LSTM deep learning model according to an embodiment of the present invention. Figure 5 shows a flowchart of a GPS path prediction system according to an embodiment of the present invention. Figure 6 shows the system performance (pedestrian data set) according to the look_back length according to an embodiment of the present invention. Figure 7 shows system performance (taxi data set) according to look_back length according to an embodiment of the present invention. Figure 8 shows the system performance according to the look_forward length according to an embodiment of the present invention. The present invention will be described in detail below with reference to the contents described in the attached drawings. However, the present invention is not limited or restricted by exemplary embodiments. Identical reference numerals in each drawing indicate components that perform substantially the same function. The purpose and effects of the present invention may be naturally understood or become clearer through the following description, and the purpose and effects of the present invention are not limited solely to the description below. Furthermore, in describing the present invention, if it is determined that a detailed description of known technology related to the present invention may unnecessarily obscure the essence of the present invention, such detailed des