Accurately recognizing vehicle driving intentions is crucial for autonomous driving. To address the issues of limited perception capabilities in complex traffic scenarios with single-vehicle intelligence, this paper proposes a vehicle driving intention recognition method based on Vehicle-Road-Cloud collaborative perception. First, an overall framework for information exchange is established through Vehicle-Road-Cloud collaborative perception, analyzing the information flow of vehicle-to-vehicle, vehicle-to-road, and road-to-cloud communication. Next, a vehicle intention recognition model is developed by combining Bi-LSTM and the XGBoost algorithm. By integrating the vehicle’s historical trajectory data with the dynamic features of surrounding vehicles, the model enhances the accuracy of driving intention recognition. Finally, the innovative Bi-LSTM bidirectional sequence processing mechanism is introduced, allowing the model to simultaneously capture both forward and backward temporal dependencies, optimizing data processing and improving the model’s robustness in complex traffic scenarios. Testing on the NGSIM dataset shows that, compared to traditional XGBoost and LSTM-XGBoost models, the Bi-LSTM-XGBoost model achieves an overall recognition accuracy of 97.4% in lane-change intention recognition and the model maintains an accuracy of 97.2% under causal constraints. Through co-simulation testing with Sumo and Carla, the impact of varying vehicle numbers on the model’s recognition efficiency is analyzed, with results indicating that the model can recognize driving intentions in realtime within 100 ms. Further testing on a real-world dataset collected from a Vehicle-Road-Cloud collaborative perception system demonstrates that the model meets real-time requirements, exhibits high trajectory prediction capability, and enhances the perception and adaptability of autonomous vehicles in complex scenarios.