Autonomous Driving Object Detection With Interpretable Confidence Scores
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches for object detection in autonomous vehicles suffer from inaccurate confidence values, leading to false or missed detections, which can result in improper path planning and potential collisions.
Innovation Solution
The proposed solution involves a Machine Learning Model (MLM) that determines a confidence score for object detections, which is a bounded scalar value directly interpretable as a confidence measure. This model accounts for complex relationships between detected objects across spatial element regions, improving the accuracy of object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the maximum coverage value of a cluster is used as the confidence value, then the confidence value can be directly interpreted as a probability measure, but this approach leads to false detections or positives as the maximum coverage value may only correspond to a small portion of the image
Solution Approach 1:
The patent introduces an intermediary processing step between coverage value aggregation and confidence determination. Instead of directly using max or sum of coverage values, the system aggregates coverage values across multiple spatial element regions and applies a normalization process that mediates between the raw coverage data and the final confidence score, ensuring both interpretability and accuracy
Solution Approach 2:
The patent transforms the confidence determination from using raw coverage values (max or sum) to using normalized aggregated coverage values. This parameter change involves dividing the aggregated coverage by the number of spatial element regions or applying a normalization function, which converts the unbounded sum into a bounded probability measure while maintaining detection accuracy
2Reliability
If the sum of the coverage values of a cluster is used as the confidence value, then all detections are accounted for, but the sum is unbounded and cannot be directly interpreted as a confidence or probability measure
Solution Approach 1:
The patent applies a normalization transformation to convert the unbounded sum of coverage values into a bounded confidence score. By dividing the sum by a normalization factor (such as the maximum possible sum or the number of spatial element regions), the system maintains comprehensive accounting of all detections while producing a interpretable probability measure
Solution Approach 2:
The patent replaces the direct arithmetic sum operation with a normalized aggregation mechanism. This substitution transforms the mechanical addition of coverage values into a probabilistic interpretation through normalization, allowing the system to account for all detections while maintaining proper confidence measurement
3Ease of manufacture
If each set of bounding box coordinates is assigned to a single spatial element region, then the training is simplified, but this approach generates false detections when objects are close to each other or occupy multiple spatial element regions
Solution Approach 1:
The patent segments the object detection task across multiple spatial element regions, allowing a single object to be detected in multiple regions. Instead of forcing each bounding box to assign to one region, the system creates multiple detections for the same object in different spatial element regions, then aggregates these detections to form a comprehensive object detection with reduced false positives
Solution Approach 2:
The patent merges multiple detections of the same object from different spatial element regions into a single aggregated detection. By combining coverage values and coordinates from multiple regions that detect the same object, the system maintains training simplicity while improving detection accuracy through the aggregation process
Data Source
AI summary
In various examples, detected object data representative of locations of detected objects in a field of view may be determined. One or more clusters of the detected objects may be generated based at least in part on the locations and features of the cluster may be determined for use as inputs to a machine learning model(s). A confidence score, computed by the machine learning model(s) based at least in part on the inputs, may be received, where the confidence score may be representative of a probability that the cluster corresponds to an object depicted at least partially in the field of view. Further examples provide approaches for determining ground truth data for training object detectors, such as for determining coverage values for ground truth objects using associated shapes, and for determining soft coverage values for ground truth objects.