Scalable Tree Goal Prediction Based on Decision Interactions

The road conditions for vehicle movement in urban intersection scenarios are complex. Autonomous vehicles must be able to effectively infer the goals of other vehicles to make reasonable decisions to complete the driving task. Considering the difficulty and specificity of inferring vehicle behavior in different scenarios, it is a challenge to predict the goals of vehicles in an accurate, interpretable and scalable way. Accordingly, this paper proposes a decision interaction based extensible tree (DIET), a scalable goal prediction system. DIET extracts common semantic features from road and vehicle information in the scenario, and extends the feature space and eliminates errors arising from simple behavioral inference by analyzing the decision interactions between surrounding vehicles. Extensive comparative experiments have been conducted on the proposed DIET and several other typical related works under different scenarios. The model reported in this paper outperforms state-of-the-art methods by 10.9% in accuracy of new scene predictions. Clearly, DIET is able to more accurately predict the vehicle's goal in different scenarios, and the process is interpretable.