Vehicle Extent Correlation from Noisy Limited FOV Frames
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle detection systems face challenges in accurately correlating vehicle extents and locations, especially with noisy detections and limited field-of-view image frames, which can lead to inconsistent and costly tolling operations.
Innovation Solution
A machine-learning based approach using convolutional neural networks (CNNs) that pairs front and back vehicle detections across multiple frames, employing interpolation and scoring mechanisms to establish vehicle pools, thereby determining vehicle location and extent even with partial or obscured views.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional vehicle detection methods are used with limited field-of-view frames, then detection speed is maintained, but detection accuracy and reliability deteriorate due to noisy detections and obscured views
Solution Approach 1:
The patent transitions from analyzing single 2D image frames to processing sequences of frames over time, adding the temporal dimension. This allows the system to track vehicle fronts and backs across multiple frames, correlating detections temporally to establish vehicle pools even when individual frames have limited field-of-view or noisy detections
Solution Approach 2:
The patent merges multiple detection results from different frames by creating vehicle pools that combine front detections and back detections across time. This pooling approach consolidates noisy individual detections into more reliable vehicle extent and location estimates through temporal aggregation
2Measurement precision
If multiple image frames are processed to improve vehicle tracking, then vehicle location accuracy improves, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the detection problem into independent front detection and back detection tasks, processing them separately through the CNN. This segmentation allows for specialized processing of each vehicle end and simplifies the overall correlation logic by treating front and back detections as distinct entities to be matched
Solution Approach 2:
The patent performs preliminary correlation of front and back detections across frames before final vehicle pool assignment. By pre-identifying potential front-back pairs and their temporal relationships, the system reduces the computational complexity of final vehicle tracking and pool creation
Data Source
AI summary
A computer accesses a plurality of image frames. The computer identifies, within the plurality of image frames, a plurality of vehicle front vehicle back detections. The computer pairs at least a subset of the plurality of vehicle back detections with vehicle front detections. A given vehicle back detection is paired with a given vehicle front detection based on camera angle relative to a predefined axis. The computer assigns, using each of a plurality of pools, a score to each vehicle front detection—vehicle back detection pair, each non-paired vehicle front detection, and each non-paired vehicle back detection. Each pool comprises a data structure representing a scoring mechanism and a set of detections. The computer assigns each detection to a pool that assigned a highest score to that detection. Upon determining that a given pool comprises at least n detections: the computer labels the given pool as representing a specific vehicle.