CNN-LSTM Method for Vehicle Motion Behavior Recognition

Recognizing vehicle motion behavior is crucial for navigation and positioning. Practical recognition of vehicle behavior can effectively constrain the noise parameters of the Kalman filtering algorithm, thereby enhancing navigation and positioning precision. This paper proposes using a CNN-LSTM combined neural network to identify vehicle motion behavior, given the low recognition accuracy of traditional physical modeling methods and the issue that machine learning must rely on manual experience to extract features. The accelerometer, gyroscope, and geomagnetic sensor are used to capture data, followed by Savitzky-Golay filtering and normalization of the raw data. The CNN neural network is then used to extract the spatial characteristics of the motion behavior, while the LSTM neural network is employed to extract the time series characteristics. The model structure of a combined CNN-LSTM neural network is investigated. Finally, the combined model is utilized to determine the motion behaviors of vehicles. Experiments are conducted to demonstrate that the method can identify vehicle motion behaviors effectively.