Driver Intent Prediction Using Lane and Trajectory Encoders
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for predicting driver intent in autonomous vehicles are limited to simple intents and struggle to accurately account for complex intents, lane networks, traffic signals, and past driving trajectories, leading to unstable predictions.
Innovation Solution
A method and apparatus for predicting driver intent using a grid map, which involves receiving driving environment and information, extracting lane and trajectory feature vectors using encoders, and inputting these vectors into a driver intent prediction model to predict complex driver intents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods are used to predict driver intent based on limited information such as vehicle speed, steering angle, and yaw rate, then the prediction process is simple and fast, but the prediction accuracy is insufficient for complex driver intents such as sudden lane changes, joining a lane, and biased driving within a lane
Solution Approach 1:
The prediction model is segmented into multiple independent encoders: a lane information encoder that processes lane network data, a trajectory encoder that processes historical trajectory data, and a driving information encoder that processes vehicle state data. Each encoder independently extracts features from its specific input type, and their outputs are concatenated to form a comprehensive driver intent prediction. This segmentation allows the system to handle complex intents by combining specialized feature extraction capabilities without creating an overly complex monolithic model.
Solution Approach 2:
The invention introduces a new dimension of lane network information representation by encoding lane connectivity, intersections, and road structure into feature vectors. This adds spatial and topological context that conventional methods lacking lane network data cannot capture, enabling the system to distinguish between similar driving scenarios based on road structure (e.g., approaching an intersection vs. open road) and predict intents like lane joining or biased driving that are causally related to lane network features.
2Productivity
If conventional methods ignore lane network, traffic signals, and past driving trajectories to simplify the prediction process, then the computational load is reduced, but the ability to accurately predict complex driver intents is compromised
Solution Approach 1:
The prediction model is segmented into multiple independent encoders: a lane information encoder that processes lane network data, a trajectory encoder that processes historical trajectory data, and a driving information encoder that processes vehicle state data. Each encoder independently extracts features from its specific input type, and their outputs are concatenated to form a comprehensive driver intent prediction. This segmentation allows the system to handle complex intents by combining specialized feature extraction capabilities without creating an overly complex monolithic model.
Solution Approach 2:
The system performs preliminary encoding of lane information and trajectory information into fixed-dimensional feature vectors before the main prediction process. The lane information encoder pre-processes lane network data into a lane feature vector, and the trajectory encoder pre-processes historical trajectories into a trajectory feature vector. These pre-computed feature vectors are then combined with driving information for the final prediction, reducing the computational burden during real-time prediction while preserving rich contextual information.
3Adaptability or versatility
If artificial intelligence methods are used to predict destination coordinates for future route prediction, then the system can handle complex scenarios, but the prediction results become unstable due to inputs not considered during training
Solution Approach 1:
The system applies local quality by designing specialized encoders for specific input types: the lane information encoder focuses exclusively on extracting spatial and topological features from lane network data, the trajectory encoder focuses on temporal patterns from historical trajectories, and the driving information encoder focuses on vehicle state features. Each encoder is optimized for its specific data type, ensuring that the model considers all relevant input types during training and deployment, thereby improving reliability and reducing instability in predictions for complex scenarios.
Data Source
AI summary
An embodiment method performed by a driver intent prediction apparatus for predicting a driver intent of a driver includes receiving driving environment information including lane information and trajectory information, receiving driving information, outputting a lane feature vector by inputting the lane information into a first encoder, outputting a trajectory feature vector by inputting the trajectory information into a second encoder, and predicting the driver intent by inputting a plurality of feature vectors into a driver intent prediction model.