Autonomous Driving Vision Feedback for Critical Object Verification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vision systems for motor vehicles face challenges in efficiently utilizing processing resources to achieve reliable intervention decisions while minimizing false interventions, often requiring a trade-off between computational cost and functional performance.
Innovation Solution
A feedback loop is established between a decision module and a critical object classifier, prioritizing critical objects by identifying those with high collision probability and allocating processing resources to verify their classification, allowing for optimal resource utilization by splitting the processing time budget between an object detector and a more powerful critical object classifier.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more computationally expensive features and classifiers are used to improve object detection performance, then functional performance is improved, but the system becomes slower and more expensive
Solution Approach 1:
The system segments object detection into two stages: a first object detector performs rapid initial detection with lower computational cost, and a second critical object classifier performs detailed classification only on detected objects. This segmentation allows the system to achieve high detection accuracy through the second classifier while maintaining processing speed by limiting its application scope.
Solution Approach 2:
The system applies different levels of classification quality to different objects. The second critical object classifier provides high-quality, computationally expensive classification only for objects that pass the first detector's screening, rather than applying uniform high-quality classification to all potential objects. This local application of high-quality processing maintains overall system speed while ensuring accurate classification where needed.
2Reliability
If a powerful classifier is applied to all detected objects to ensure high classification reliability, then classification accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The first object detector performs preliminary detection and filtering before the second critical object classifier is applied. By pre-identifying and filtering relevant objects in the first stage, the system prepares a reduced set of candidates for the computationally expensive second classifier, thereby reducing overall processing time while maintaining high classification reliability for critical objects.
Solution Approach 2:
The system applies the powerful second critical object classifier partially, only to objects that are deemed relevant by the first detector, rather than applying it excessively to all possible objects in the scene. This partial application significantly reduces processing time while maintaining high classification reliability for the subset of objects that require such detailed analysis.
3Productivity
If the object detector uses the entire processing time budget to detect all objects, then detection completeness improves, but time for verification of critical objects is reduced
Solution Approach 1:
The second critical object classifier operates in a feedback loop with the first object detector. Objects detected by the first detector are fed into the second classifier for verification, and the results can inform subsequent detection cycles. This feedback mechanism ensures that critical objects receive thorough verification while the system maintains efficient overall detection throughput through the first detector's rapid initial screening.
Data Source
Figure 1~2
AI summary
A vision system and method for autonomous driving and/or driver assistance in a motor vehicle are presented, comprising an imaging apparatus (11) adapted to capture images (30) from a region surrounding the motor vehicle, a data processing device (14) comprising an object detector (31) adapted to detect and classify objects in the vehicle surrounding by processing images captured by said imaging apparatus (11), and a decision module (32) adapted to estimate a collision probability of a detected object with the motor vehicle, and to make an intervention decision (37) in case a detected object has a non-tolerable collision probability. A critical object classifier (35) is connected to the decision module (33) in a feedback loop (36). The decision module (33) is adapted to identify, among all detected objects, a sub-group of critical objects having a high potential of causing an intervention, and to feed information regarding identified critical objects to the critical object classifier (35). The critical object classifier (35) is adapted to verify the classification of the object detector (31), and to feed back a result of the verification to the decision module (33).