With the development of sensor technology, context-aware vehicles (e.g., location services and automatic driving) are becoming increasingly popular. However, these applications require high context perception accuracy, especially in assisted and automatic driving, which has increasingly high requirements for the continuity, reliability, and accuracy of vehicle positioning. The positioning performance of a single Global Positioning System (GPS) may be decreased by various factors, such as occlusion and interference when driving on urban roads [
1,
2]. Thus, it is difficult to meet the needs of the Internet of Vehicles. However, multi-source sensor fusion, e.g., Inertial Navigation System (INS) and GPS integrated navigation systems, can effectively solve these problems [
3,
4].
For the data fusion problem in integrated navigation, Kalman Filter (KF) is the existing optimal trajectory estimation method, which solves the problem of tedious calculation caused by the weak nonlinear ability. Particle Filter (PF) is considered a benchmark of the filtering method for predicting vehicle position, but the large number of particles required by PF leads the algorithm computationally expensive. In integrated navigation, when the GPS signal is interrupted [
5,
6], the positioning error in the Inertial Measurement Unit (IMU) accumulates over time, decreasing the overall performance of the integrated navigation [
7,
8,
9]. To improve the positioning performance during GPS signal interruption, Artificial Neural Networks (ANNs) have been introduced into the INS/GPS integrated navigation system, e.g., Multilayer Perceptron Neural Networks (MLPNNs) [
10,
11], Radial Basis Function Neural Networks (RBFNNs) [
12,
13,
14], Long Short Term Memory Recurrent Neural (LSTM-RNNs) [
15], networks and adaptive Neuron-Fuzzy Inference Systems (ANFISs) [
16,
17]. The main idea is to train the relationship between the vehicle feature data and INS errors through ANNs when the GPS signal is available. When the GPS signals are available, the current or latest training model is used to predict the positioning data. This method effectively reduces the positioning error and ensures positioning continuity. Since the ANN is trained completely using input data, its generalization ability is limited when the vehicle state data during training is different from that during prediction. To solve this problem, in one study [
18], ensemble learning was included in the INS/GPS integrated navigation system, which effectively improved the generalization ability. Although the Least Squares Boosting and Bagging algorithms proposed in that study [
18] could improve the positioning accuracy, the errors of the INS internal sensor (e.g., steering deviation, running deviation, and scale factor drift) increased the nonlinear complexity of the relationship between the input and output data. The model was weak at recognizing feature variables, resulting in unsatisfactory predictions of positioning. In addition, the sensor’s errors accumulated over time. When the GPS signal loss was over 5 min, the prediction accuracy of the ensemble learning scheme began to decrease gradually [
18].
To solve the above-mentioned problems, a Kalman Filter-Gradient Boosting Decision Tree-Particle Swarm Optimization (KF-GBDT-PSO, henceforth denoted KGP) data fusion method is proposed herein [
19], which consists of two consecutive phases: Training and prediction. In the training phase, the KGP prediction model can compensate for the INS positioning error through the relationship between vehicle feature data and KF estimations of the positioning error [
20,
21]. In the prediction phase (during GPS signal loss), the trained model immediately predicts the positioning data. Compared with ANNs, the predicted values of the GBDT are obtained through accumulating the residuals of multiple trees. Due to its advantage of reducing model deviations, the Gradient Boosting Decision Tree (GBDT) provides a better generalization ability with better accuracy. Additionally, selecting regression parameters can be challenging. Thus, Particle Swarm Optimization (PSO) is introduced in the training phrase to select the optimal parameters for GBDT [
22]. The KGP not only could extract nonlinear vehicle feature data in parking and driving states using the addition model and the forward distribution algorithm but could also use the Huber loss function to eliminate the location outliers collected due to road complexity. By flexibly covering various types of variables, the error rate of the integrated navigation system was effectively reduced [
23], increasing the prediction accuracy of the positioning.
The remainder of this paper is organized as follows. In
Section 2, an overview of the GBDT and PSO theories is provided. In
Section 3, the integration scheme for the KGP method is introduced and discussed in detail. The experimental results are discussed in
Section 4. Conclusions are presented in
Section 5.