Camera Pose Alignment for Rotated-Object Target Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current target detection methods require data enhancement through geometric transformations on neural networks to handle rotated images, leading to increased computing time and resource consumption.

Innovation Solution

Adjust the captured scene image based on camera position and pose information to align the image orientation with the actual orientation, allowing direct target detection without data enhancement, using SLAM systems to determine rotation angles and perform target detection on adjusted images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If data enhancement through geometric transformations is performed on neural networks to handle rotated images, then the neural network can learn characteristics of rotated objects, but computing time and computing resources are increased

Engineering Contradiction:

Improveability to detect rotated objectsVSAvoidcomputing time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

Instead of transforming training data to match various camera orientations, the patent inverts the approach by transforming the captured image to match the fixed training data orientation. The image is rotated based on camera pose information to align with the horizontal orientation expected by the neural network, eliminating the need for complex data enhancement while maintaining detection accuracy for rotated objects

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent performs image orientation adjustment as a preliminary step before target detection. By pre-aligning the captured image to the horizontal orientation using camera pose information from SLAM, the system prepares the image in advance to match the training data orientation, allowing direct use of the trained neural network without additional transformation during detection

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

Engineering Contradiction:

Improveability to detect rotated objectsVSAvoidcomplexity of data enhancement process

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent simplifies the system by inverting the transformation approach: rather than creating multiple transformed versions of training data, it transforms the single captured image to match the fixed training orientation. This eliminates the complexity of data enhancement pipelines while maintaining the ability to detect objects in any orientation

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent introduces camera pose information from SLAM as an intermediary that bridges the captured image and the fixed training data orientation. This intermediary provides the rotation angle needed to align the image, replacing the need for complex data enhancement mechanisms while maintaining adaptability to rotated objects

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentEP4073690B1Target detection method, terminal device, and medium

Publication Date: 2025.10.29 GUANGDONG OPPO MOBILE TELECOMMUNICATIONS CORP LTD

Figure 1~2
Figure 3~4
Figure 5~6

AI summary

The present disclosure provides a target detection method. The method includes: (101) acquiring a first scene image captured by a camera; (102) acquiring current position and pose information of the camera; (103) adjusting the first scene image based on the current position and pose information of the camera to obtain a second scene image; (104) and performing a target detection on the second scene image. In addition, The present disclosure also provides a terminal device, and a medium.