Object Tracking in Non-Linear Distortion Lenses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for object tracking across multiple video cameras with non-linear distortion lenses, such as wide-angle and fisheye lenses, face challenges in accurately mapping feature points due to distortion, leading to misjudgment and failure in tracking objects, especially when cameras have different viewing angles and large distances between them.
Innovation Solution
A deep learning method is employed to train three networks: an object identification model, an object comparison model, and a coordinate mapping model. These models identify objects, determine corresponding objects across multiple cameras, and establish position correspondences, allowing for accurate object tracking across non-linearly distorted images without relying on specific distortion models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional feature point methods are used for object tracking, then tracking accuracy is improved in linear distortion scenarios, but tracking reliability deteriorates when non-linear distortion lenses are used
Solution Approach 1:
The patent transforms the approach from directly matching feature points in distorted images to first identifying objects, then using object centroids as reference points for coordinate mapping. This parameter transformation (from feature point coordinates to object centroid coordinates) resolves the contradiction by making the tracking system reliable under non-linear distortion while maintaining accuracy through deep learning-based object identification
Solution Approach 2:
The patent introduces object centroids as an intermediary element between the distorted image coordinates and the tracking system. Instead of directly using distorted feature points, the system uses object centroids (obtained through deep learning object identification) as mediators to establish coordinate mapping relationships, thereby ensuring both reliability and accuracy in non-linear distortion scenarios
2Reliability
If deep learning models are used for object identification and coordinate mapping, then tracking reliability is improved in non-linear distortion scenarios, but computational complexity increases
Solution Approach 1:
The patent segments the complex tracking problem into three distinct deep learning models: object identification model, object comparison model, and coordinate mapping model. Each model handles a specific sub-task, which improves reliability through specialized processing while managing computational complexity by distributing the workload across multiple focused models rather than one monolithic complex system
Solution Approach 2:
The patent extracts and separates the coordinate mapping function from the object identification function. The coordinate mapping model takes object centroids (extracted from identified objects) as input and independently handles the mapping to different camera coordinate systems. This extraction improves reliability by dedicating specific models to specific tasks while optimizing computational efficiency through functional separation
3Measurement precision
If manual or semi-automatic distortion measurement is performed using checkerboard images, then measurement precision is improved, but ease of operation deteriorates
Solution Approach 1:
The patent implements a self-service approach where the system automatically performs distortion compensation and coordinate mapping without requiring manual checkerboard calibration. The deep learning models automatically adapt to the distortion characteristics of each lens type, eliminating the need for manual measurement operations while maintaining high precision through automated learning from training data
Solution Approach 2:
The patent performs preliminary distortion characterization during the training phase of the deep learning models. Instead of requiring distortion measurement at operation time, the models are pre-trained with distortion characteristics, so the distortion compensation is already built into the model parameters. This preliminary action eliminates manual operation requirements while maintaining measurement precision
Data Source
AI summary
An object tracking method and system in multiple non-linear distortion lenses are provided. A deep learning method is used for training an object identification model, an object comparison model, and a coordinate mapping model. The object identification model and the object comparison model are used for identifying and comparing objects with non-linear distortion respectively in order to find a plurality of groups of corresponding object information for the same objects in visions of multiple image capturing apparatuses. The coordinate mapping model is used for verifying the plurality of groups of corresponding object information, and finding all position mappings in visions of multiple image capturing apparatuses via the verified plurality of groups of corresponding object information.