Autonomous Driving Speed Control for Unsignalized Intersections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous driving technologies face challenges in safely navigating intersections without traffic lights, often prioritizing safety over efficiency, leading to traffic congestion and requiring costly infrastructure for vehicle communication, and neural network methods are not effective alternatives.
Innovation Solution
An autonomous driving device employing a deep reinforcement learning system with spatial and temporal attention networks to determine a target speed based on a history of state information, including location, driving state, and trajectory data, using a spatial attention network to assess spatial importance and a temporal attention network to apply temporal importance for efficient and safe vehicle control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conservative control is applied to avoid collisions, then safety is improved, but traffic congestion increases
Solution Approach 1:
The system dynamically adjusts the target speed based on real-time spatial and temporal attention weights, allowing the vehicle to switch between conservative and aggressive driving strategies depending on the current situation. The execution network continuously updates the speed command based on changing environmental conditions, resolving the contradiction between safety and efficiency.
Solution Approach 2:
The system changes the parameters of driving behavior by using attention mechanisms that dynamically weight different state information. The spatial attention network adjusts the importance of different spatial locations, while the temporal attention network adjusts the importance of different time steps, allowing the system to adapt its safety vs. efficiency balance through parameter changes rather than fixed conservative rules.
2Reliability
If vehicle communication infrastructure is deployed to share information, then passing time scheduling is improved, but infrastructure cost increases
Solution Approach 1:
Each autonomous vehicle independently performs spatial and temporal attention analysis on its own sensor data and state history. The system processes and interprets environmental information locally without requiring external communication infrastructure, allowing vehicles to self-determine safe passing times based on their own observations and learned patterns.
Solution Approach 2:
The patent extracts the essential function of vehicle-to-vehicle communication by implementing the attention mechanism that processes only the most relevant spatial and temporal information locally. This eliminates the need for comprehensive communication infrastructure while retaining the scheduling benefits through selective information processing.
3Ease of manufacture
If conventional neural network supervised learning is used, then training is simplified, but effectiveness at intersections is insufficient
Solution Approach 1:
The system performs preliminary action by pre-training the execution network on comprehensive driving data including intersection scenarios, then further fine-tunes it through reinforcement learning with spatial and temporal attention mechanisms. This two-stage approach simplifies the overall training process while achieving high effectiveness at intersections.
Solution Approach 2:
The system implements feedback through reinforcement learning where the execution network receives rewards or penalties based on its driving decisions at intersections. This feedback loop continuously improves the network's performance in intersection scenarios while maintaining the simplicity of the underlying neural network architecture.
Data Source
AI summary
An autonomous driving device includes an execution network configured to determine a target speed of a driving vehicle according to state information history including a plurality of state information for road environment. The plurality of state information are generated at a plurality of times. The execution network includes a spatial attention network configured to receive the state information history and to generate feature data reflecting spatial importance based on the state information history; and a temporal attention network configured to determine the target speed of the driving vehicle by applying temporal importance to an output of the spatial attention network.