一种面向端到端规控的无人物流车感知特征输出方法及系统
By using multi-scale feature fusion and implicit feature encoding, the problems of information loss and non-differentiability in the autonomous driving perception module are solved, enabling end-to-end joint optimization and safe decision-making, thereby improving the decision-making performance and safety of the autonomous driving system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HONEYCOMB (WUHAN) MICROSYSTEM TECH CO LTD
- Filing Date
- 2026-05-13
- Publication Date
- 2026-07-17
AI Technical Summary
The discrete output of existing autonomous driving perception modules results in severe information loss, making it impossible to achieve gradient backpropagation and uncertainty transmission, thus hindering risk perception for end-to-end joint optimization and regulatory decision-making.
By employing multi-scale feature fusion and implicit feature encoding, and outputting BEV features and implicit feature vectors, combined with an uncertainty estimation head, a differentiable interface connection between the perception module and the planning and control module is achieved, ensuring gradient propagation and uncertainty assessment, and switching to a conservative planner to ensure safety.
It preserves the semantic, spatial, and motion details of the scene, enables end-to-end joint training, improves decision-making performance, and switches to conservative planning under high uncertainty to ensure safety and efficiency.
Smart Images
Figure CN122172772B_ABST