Machine Learning for Autonomous Vehicle Behavior Prediction in Mixed Traffic Conditions

Abstract

The great mass of road users and the variation of their behavior patterns make defining, classifying, and recognizing driving intentions a challenging task. Prediction of other vehicles' future behavior is crucial for the automated cars in mixed traffic conditions to ensure the progression of driving smoothly and safely. In the real-world data, an observed driving behavior like the position, velocity or acceleration of the vehicles, crosses with finite state machines (FSMs), are included in a simple and perform able manner to evaluate its performance on the prediction of other vehicles' end state. Variance layers of LSTM networks were used to exploit the sequence patterns in the driving data, and also to form a supervised prediction model to seize the next behavior of the vehicle in a real-time operation. On the highway, the CACC-ready vehicles passed 23,558 times and collected a total of 195,150.87 m transportation data. The driving behavior prediction model achieves an average prediction error (in the estimation of end state) is equal 3.24 note meters or 0.04 s.