Multi-Modal Identity Recognition for Occlusion and Large-Angle Scenes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing identity recognition technologies face challenges with low accuracy due to issues such as large angles, occlusions, and changes over time, particularly in complex scenes, leading to missed and false recognitions, and pedestrian re-identification fails to accurately determine personnel identity despite robustness against short-term occlusions.
Innovation Solution
An identity recognition method utilizing multi-modal features, including face and human body features, is employed, with a pre-established feature extraction model to compare these features against a registry, determining the highest comparison score to identify the target object accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If face recognition based on CNN and cosine function is used, then identity recognition can be performed, but accuracy decreases due to large angles, occlusions, and changes over time
Solution Approach 1:
The patent combines face recognition and ReID into a unified multi-modal recognition system. The feature extraction model simultaneously extracts face features and body features, which are then fused into comprehensive feature vectors for comparison against the registration library. This merging allows the system to leverage both modalities' strengths: face features for high-accuracy identification and body features for robustness against occlusions and pose variations.
Solution Approach 2:
The patent creates composite feature vectors by fusing face features and body features. The loss function is designed as a composite of face loss and body loss, where each modality contributes to the overall feature representation. This composite approach enables the system to maintain high accuracy when face features are clear while providing robustness through body features when faces are occluded or at large angles.
2Reliability
If ReID based on human body information is used, then robustness against short-term occlusion and large angles is improved, but accuracy decreases due to inability to cope with dressing transformation
Solution Approach 1:
The patent merges ReID's body feature extraction capability with face recognition. The dual-branch feature extraction model processes both face and body regions simultaneously, creating complementary feature representations. This combination allows the system to use body features for robust tracking under occlusion while using face features for accurate identity determination, resolving the trade-off between robustness and precision.
3Productivity
If single-modal face feature extraction is used, then processing speed is maintained, but accuracy decreases in complex scenes with occlusions and large angles
Solution Approach 1:
The patent segments the recognition process into parallel face feature extraction and body feature extraction branches. This segmentation allows independent optimization of each modality's processing pipeline while maintaining overall system efficiency. The feature fusion occurs at the vector level, preserving computational speed while enhancing accuracy through multi-modal information integration.
Data Source
AI summary
An identity recognition method, a computer apparatus, a non-transitory computer-readable storage medium are provided. The method includes: acquiring a to-be-identified image including a test object; extracting multi-modal features of the test object from the to-be-identified image based on a pre-established feature extraction model, the multi-modal features including at least one face feature and one human body feature; comparing the multi-modal features to modal features included by multi-modal feature sets in a pre-established feature registry; determining a target multi-modal set corresponding to the highest comparison score from the multi-modal feature sets, wherein each of the multi-modal feature sets incorporates at least one of the face feature and the human body feature; and determining an identity information corresponding to a target face feature included by the target multi-modality set, and determining the identity information corresponding to the target face feature as an identity information of the test object.