Vehicle Object Recognition Using Posterior Confidence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition techniques are prone to erroneous recognition of objects, either identifying non-vehicles as vehicles or vice versa, especially when a vehicle's windshield has a water-repellent coating and the wiper is not used in rainy conditions, leading to increased likelihood of misclassification.
Innovation Solution
An object recognition device equipped with a radar unit, characteristic determination unit, cluster forming unit, and tracking processing unit that uses the results of multiple time points to determine whether an object is a vehicle or not, calculating posterior confidence based on vehicle and non-vehicle characteristics to accurately associate and recognize clusters, thereby reducing erroneous recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object recognition is performed based on reflection points from transmission waves, then object detection capability is improved, but erroneous recognition of non-vehicles as vehicles increases
Solution Approach 1:
The patent segments the recognition process into multiple independent stages: cluster formation from reflection points, tracking target association, and sequential probability calculation. Each stage processes specific aspects of object identification separately, allowing comprehensive evaluation while maintaining detection sensitivity. This segmentation enables the system to distinguish vehicles from non-vehicles more reliably by evaluating multiple criteria in sequence.
Solution Approach 2:
The patent performs preliminary clustering of reflection points into candidate groups before final object recognition. By pre-organizing reflection points into clusters based on spatial and temporal relationships, the system prepares structured data that facilitates more accurate subsequent recognition. This preliminary action reduces erroneous recognition by establishing probable object boundaries before classification.
2Productivity
If association is performed based on single time point data, then processing speed is improved, but recognition accuracy deteriorates
Solution Approach 1:
The patent implements periodic association processing where tracking targets are associated with clusters at multiple discrete time points. The system calculates association probabilities sequentially across different measurement cycles, updating recognition confidence as new data becomes available. This periodic approach maintains processing efficiency while improving accuracy through temporal accumulation of evidence.
Solution Approach 2:
The patent employs feedback mechanisms where recognition results from previous time points inform subsequent association decisions. The system calculates posterior probabilities that incorporate previous recognition outcomes, creating a feedback loop that continuously refines object identification. This feedback mechanism enables accurate recognition while maintaining efficient processing by building upon previous computational results.
3Reliability
If multiple time points are used for object recognition, then recognition accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent implements dynamic processing where the system adapts its computational approach based on the number of tracking targets and clusters. The association processing unit dynamically determines which tracking targets require association with which clusters, adjusting computational effort to actual scene complexity. This dynamic approach maintains high recognition accuracy while minimizing unnecessary processing of already-identified objects.
Solution Approach 2:
The patent changes processing parameters based on measurement cycle progression and object confidence levels. The system adjusts association thresholds, probability calculation depths, and processing priorities according to the current recognition state. This parameter adaptation enables accurate multi-time-point recognition while managing computational complexity by focusing resources on uncertain or newly-detected objects.
Data Source
AI summary
An object recognition device mountable to a vehicle includes a tracking processing unit including a first association processing unit, a second association processing unit, and a vehicle likelihood calculation unit. The first association processing unit extracts, for each tracking target and based on a positional relationship, a provisionally associated cluster which is a cluster assumed to indicate the same target as the tracking target. The second association processing unit calculates a posterior confidence indicating a probability that when it is assumed that the provisionally associated cluster and the tracking target indicate the same target, the target is a vehicle, determines whether to perform association based on the calculated posterior confidence, and calculates the posterior confidence based on the posterior confidence calculated for the tracking target in the previous measurement cycle and on a cluster vehicle likelihood of the provisionally associated cluster calculated by the vehicle likelihood calculation unit.