Convolutional Neural Network Maneuver Prediction via Frenet Space
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current advanced driver assistance systems (ADAS) have poor prediction performance for other traffic participants' maneuvers, leading to inadequate longitudinal and lateral control, delayed detection of traffic anomalies, and safety risks due to independent prediction methods that do not consider surrounding vehicles' interactions.
Innovation Solution
A training method for a convolutional neural network that transforms traffic data into Frenet space, incorporating additional traffic parameters such as social interactions, road geometry, and environmental data, enabling the network to predict maneuvers and detect anomalies by considering the interactions of all traffic participants, rather than relying solely on vehicle dynamics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If independent prediction of each traffic participant is used based on their sensory inputs only, then the prediction process is simple and computationally efficient, but the prediction performance is poor and does not capture social interactions between vehicles
Solution Approach 1:
The patent merges independent vehicle predictions with social interaction analysis by integrating a social pooling unit that processes relationships between multiple traffic participants. This combines individual vehicle dynamics models with a collective social behavior model, allowing the system to capture interactions between vehicles while maintaining the computational efficiency of independent predictions.
Solution Approach 2:
The prediction system is designed to perform multiple functions: it independently predicts each vehicle's maneuver based on its own sensory inputs, simultaneously analyzes social interactions between all traffic participants, and integrates these predictions to produce a comprehensive forecast. This multi-functional approach improves prediction reliability without requiring separate specialized systems.
2Reliability
If rule-based approaches are used for maneuver prediction considering surrounding vehicles, then social interactions are partially captured, but the approach lacks generality and cannot detect traffic anomalies early
Solution Approach 1:
The patent replaces rule-based mechanical decision-making with a neural network-based social pooling unit. This unit learns social interaction patterns from data and can generalize to unseen scenarios, enabling early anomaly detection through probabilistic modeling rather than fixed rules. The system can identify unusual traffic patterns that deviate from learned normal behavior.
Solution Approach 2:
The system changes from deterministic rule-based parameters to probabilistic parameters that capture the uncertainty and variability in traffic participant behavior. The social pooling unit outputs probability distributions over possible maneuvers, allowing the system to detect anomalies when observed behavior significantly deviates from expected probability distributions.
3Measurement precision
If Frenet space transformation is applied to traffic data, then the convolutional neural network can effectively process spatial relationships and improve prediction accuracy, but the data processing complexity increases
Solution Approach 1:
The patent applies Frenet space transformation to convert Cartesian coordinates into a curvilinear coordinate system that follows the road geometry. This transformation accounts for the curved nature of roads and preserves spatial relationships between vehicles on multi-lane roads, enabling the convolutional neural network to accurately process spatial patterns regardless of road curvature or vehicle orientation.
4Reliability
If additional traffic parameters beyond position and lane assignment are incorporated, then the prediction model captures more comprehensive traffic scenario information, but the computational load and model complexity increase
Solution Approach 1:
The convolutional neural network is designed to process multiple types of input data (positions, lane assignments, velocities, accelerations, and other traffic parameters) through a unified architecture. This multi-functional design allows the system to comprehensively analyze traffic scenarios using diverse parameters while managing computational load through efficient network operations and shared feature extraction.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The invention relates to a training method for a convolutional neural network for predicting a driving maneuver of at least one traffic participant (V0, V1) in a traffic scenario (S) of an ego-vehicle (E), comprising the steps: determining (S1) traffic data for the convolutional neural network (10), wherein the traffic data comprises a position of each of the traffic participants (V0, V1), a lane assignment for each of the traffic participants (V0, V1) and at least one additional traffic parameter, wherein the at least one additional traffic parameter relates to the traffic scenario (S); wherein determining (S2) transformed traffic data by transforming positional coordinates of the traffic data into Frenet space; determining (S3) a trained convolutional neural network (10) by training the convolutional neural network dependent on the determined transformed traffic data.