Multi-Layer Video Object Detection for Precise Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer vision technologies face challenges in accurately detecting and tracking target objects in video sequences due to limitations in feature extraction and bounding box estimation, leading to suboptimal tracking performance.
Innovation Solution
A method and device utilizing a neural network with multiple layers to extract features from reference and detection frames, combining classification and regression results from different network layers to determine precise bounding boxes for target objects, enhancing tracking accuracy and speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single network layer is used for feature extraction, then the device complexity is reduced, but the measurement precision of target object detection deteriorates
Solution Approach 1:
The patent divides the feature extraction process into multiple network layers, where different layers extract features at different levels of abstraction. Each layer processes features independently and contributes to the final detection result, thereby improving detection precision while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent introduces a temporal dimension by incorporating reference frames from previous time steps alongside current frame features. This multi-dimensional feature fusion (spatial features from multiple layers + temporal features from reference frames) enhances detection precision without proportionally increasing complexity.
2Measurement precision
If multiple network layers are used for feature extraction, then the detection precision is improved, but the processing speed deteriorates
Solution Approach 1:
The patent performs feature extraction from reference frames in advance and stores these features for rapid retrieval during detection. This preliminary processing reduces the computational burden during real-time detection, allowing multiple network layers to be used without proportionally increasing processing time.
Solution Approach 2:
The patent selectively applies multiple network layers only when necessary for accurate detection, rather than always using the full depth of the network. This partial application of complex processing maintains precision for difficult cases while preserving speed for simpler detection scenarios.
3Reliability
If features from multiple network layers are combined, then the tracking accuracy is improved, but the computational complexity increases
Solution Approach 1:
The patent applies different processing strategies to features from different network layers based on their specific characteristics. Shallow layers with high spatial resolution are used for precise localization, while deeper layers with high semantic information are used for classification, optimizing the use of each feature type without unnecessary computational overhead.
Solution Approach 2:
The patent uses reference frame features as templates or copies that are compared against current frame features. This copying approach simplifies the combination process by establishing clear correspondence relationships between temporal features, reducing the complexity of fusing multi-layer features across time steps.
4Speed
If the reference frame image size is reduced, then the processing speed is improved, but the information completeness deteriorates
Solution Approach 1:
The patent compensates for the reduced spatial dimensions of the reference frame by enriching the temporal and semantic dimensions. Features from multiple network layers provide additional channels of information, ensuring that reducing spatial resolution does not result in overall information loss.
Solution Approach 2:
The patent creates a composite feature representation by combining downscaled reference frame features with features from multiple network layers. This composite approach maintains processing speed through downsampling while preserving information completeness through multi-layer feature enrichment.
Data Source
AI summary
A method of detecting target object includes: extracting, through a neural network, a feature of a reference frame and a feature of a frame under detection; inputting each of at least two feature groups from at least two network layers of the neural network into a detector so as to obtain a corresponding detection result group output from the detector; wherein each feature group includes features of the reference frame and of the frame under detection, each detection result group includes a classification result and a regression result with respect to each of a plurality of candidate boxes for a feature group; and acquiring a bounding box for the target object in the frame under detection according to the at least two detection result groups.