Bayesian Fused Object Tracking Under Sensor Occlusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches to sensor fusion in autonomous driving and advanced driver assistance systems fail to adequately reduce false positive and false negative detection rates, leading to inaccurate object detection and tracking.
Innovation Solution
The method estimates and utilizes a confidence or probability of existence for each fused track by considering false positive and false negative rates of individual sensors, applying a Bayesian filtering algorithm to determine track confidence values, and excluding occluded sensor measurements, thereby improving filtering of false positives and avoiding false negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional sensor fusion approaches are used (designating primary sensor, multiple sensor confirmation, filtering input data), then object detection can be performed, but false positive and false negative detection rates remain high
Solution Approach 1:
The patent changes the parameter of confidence estimation by introducing a Bayesian statistical framework that dynamically calculates track confidence values based on sensor-specific false positive and false negative rates. This allows the system to adaptively weight sensor measurements according to their reliability characteristics, resolving the contradiction by transforming static sensor fusion into dynamic confidence-based fusion that simultaneously improves detection reliability and reduces false alarm rates
Solution Approach 2:
The patent introduces an intermediary confidence estimation mechanism that mediates between raw sensor measurements and final object detection decisions. By inserting a Bayesian filtering layer that computes track confidence values based on sensor performance characteristics, the system creates a buffer that filters out unreliable detections while preserving true objects, thereby reducing both false positives and false negatives without sacrificing detection capability
2Reliability
If multiple sensors are used to confirm object detection, then false positives can be reduced, but false negatives increase due to occlusion and field of view limitations
Solution Approach 1:
The patent applies local quality by recognizing that different sensors have different local expertise - each sensor type (camera, radar, lidar) has specific strengths in certain conditions. The Bayesian framework assigns weights to each sensor based on its local reliability characteristics, allowing the system to rely more heavily on sensors that are less likely to produce false negatives in specific situations while still benefiting from multi-sensor fusion for false positive reduction
Solution Approach 2:
The patent introduces dynamics by making the sensor weighting adaptive rather than static. The Bayesian filter dynamically adjusts the influence of each sensor based on real-time confidence estimates that incorporate sensor performance characteristics. This allows the system to flexibly respond to occlusion and field of view limitations by automatically relying more on sensors that are likely to detect the object, thereby reducing false negatives while maintaining false positive reduction
3Reliability
If sensor measurements are filtered based on different criteria to identify clutter detections, then false positives can be reduced, but detection accuracy deteriorates
Solution Approach 1:
The patent replaces mechanical filtering approaches (hard thresholds, binary inclusion/exclusion criteria) with a statistical Bayesian framework. Instead of using rigid filters that mechanically exclude measurements based on predetermined criteria, the system uses probabilistic confidence estimation that continuously evaluates the likelihood of each detection being true. This substitution allows for more nuanced discrimination between clutter and true objects, maintaining detection accuracy while still reducing false positives
Data Source
AI summary
A sensor fusion system and method are disclosed. One or more processors are operable to receive a plurality of object detection measurements from a plurality of sensors. Each of the plurality of object detection measurements are associated with a potential object detection track. A plurality of sensor confidence values associated with each of the plurality of sensors are received. A track confidence value is determined for each of the potential object detection tracks based on the received plurality of object detection measurements and the received plurality of sensor confidence values. An object detection for a potential object detection track that has a determined track confidence value meeting a predetermined detection threshold is then determined, or confirmed, and stored in a memory for subsequent use, and is relatively unaffected by a measurement from a sensor that has a field of view that omits or is occluded with respect to the given object detection track.