Perception Object Confidence Estimation for False Positive Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating object confidence in autonomous driving systems face challenges such as high data labeling costs, long iteration cycles, and poor interpretability, leading to inaccurate and unreliable responses that can cause false positives, affecting user satisfaction.
Innovation Solution
A method using backflow data fragments from autonomous driving to construct a training dataset with true and false positive labels, employing a random forest model to estimate object confidence, allowing for efficient data labeling and reducing false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a rule-based confidence estimation method is used, then the system is simple to implement, but the accuracy of confidence estimation is low and false positive indicators cannot be effectively improved
Solution Approach 1:
The patent replaces the rule-based mechanical confidence estimation system with a deep learning-based neural network system. The neural network model automatically learns complex patterns from perception data and generates confidence estimates, substituting the simple but inaccurate rule-based approach with an intelligent system that achieves both high accuracy and automated operation.
2Measurement precision
If a deep learning-based confidence estimation method is used, then the accuracy of confidence estimation is improved, but data acquisition and labeling costs are high and the iteration cycle is long
Solution Approach 1:
The patent performs preliminary action by pre-processing perception data during the data collection phase, organizing it into the required format with necessary annotations. This preliminary preparation reduces the workload during model training and iteration, enabling faster experimentation and reduced labeling costs while maintaining high accuracy requirements.
Solution Approach 2:
The system implements self-service through automated data processing pipelines where the neural network can perform self-supervised learning on unlabeled perception data. The system automatically generates pseudo-labels and iteratively improves its confidence estimation capability without requiring extensive manual labeling, thereby reducing labeling costs and accelerating iteration cycles.
3Measurement precision
If a deep learning-based confidence estimation method is used, then the accuracy of confidence estimation is improved, but the interpretability of the model is poor
Solution Approach 1:
The patent introduces interpretability intermediaries such as attention mechanism visualizations, feature importance analysis, and confidence decomposition modules that bridge the gap between the black-box neural network and human understanding. These intermediaries translate the complex internal representations into interpretable forms while preserving the high accuracy of the deep learning model.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to the field of autonomous driving technologies, and in particular, provides a method for obtaining a confidence of a perception object, a method for autonomous driving planning and control, a readable storage medium, and a smart device, aiming to solve a problem of how to estimate a more reliable confidence of a perception object based on lower labeling costs, thereby providing an effective response basis for a planning and control module of autonomous driving. According to the invention, backflow data fragments of autonomous driving are applied and labeled with true positive labels and false positive labels to construct a training dataset, a confidence estimation model is trained based on the training dataset to obtain a trained confidence estimation model, and the trained confidence estimation model is applied to estimate a confidence of a perception object in a perception fusion result.