Video Target Detection via Key Frame Feature Affine Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing target detection methods for videos require extensive feature extraction from each frame, leading to high computation costs and low detection efficiency due to repeated calculations.
Innovation Solution
The method involves determining a flow feature map and flow field between current and previous key frames, allowing for affine transformation of key frame features to obtain image features for non-key frames without repeated feature extraction, thereby reducing computational load and improving efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If feature extraction is performed on each video frame independently, then target detection can be performed on every frame, but computation power requirements increase and detection efficiency decreases
Solution Approach 1:
The patent performs feature extraction only on key frames (I-frames) in advance, and then reuses these extracted features for non-key frames (P-frames and B-frames) through motion compensation and affine transformation. This preliminary action on key frames eliminates the need for repeated feature extraction on every frame, significantly reducing computational load while maintaining detection accuracy across all frames.
Solution Approach 2:
The patent creates copies of key frame features and transforms them to represent non-key frames. By extracting features once from key frames and then copying/transforming these features for subsequent non-key frames using motion vectors and affine transformations, the system avoids redundant feature extraction while preserving frame-specific target information.
2Ease of operation
If feature extraction is performed repeatedly on each video frame, then each frame can be processed independently, but calculation power consumption increases
Solution Approach 1:
The patent performs feature extraction as a preliminary action only on key frames before video playback or processing. These pre-extracted features are then reused for non-key frames through motion compensation, eliminating the need for repeated feature extraction and significantly reducing computation power consumption during actual video processing.
Solution Approach 2:
The patent discards the computationally expensive feature extraction step for non-key frames and recovers frame-specific features by transforming key frame features using motion vectors and affine transformations. This approach maintains the ability to process frames independently in terms of detection while avoiding redundant computation.
3Productivity
If key frame features are transformed for non-key frames, then feature extraction times are reduced, but the system must maintain key frame information and motion data
Solution Approach 1:
The patent copies key frame features and transforms them for non-key frames using motion compensation techniques. This copying approach requires storing key frame features and motion data, but the storage overhead is significantly less than performing full feature extraction on every frame, as motion vectors and affine transformation parameters are much more compact than complete feature representations.
Data Source
AI summary
A target detection method is provided for a computing device. The method includes obtaining a current frame and a previous key frame corresponding to the current frame in a video frame sequence, determining a flow feature map and a flow field between the previous key frame and the current frame, obtaining, in response to determining the current frame is a non-key frame according to the flow feature map, a key frame feature corresponding to the previous key frame, and performing affine transformation on the key frame feature according to the flow field to obtain an image feature corresponding to the current frame, and performing target detection on the current frame according to the image feature to obtain a target detection result of the current frame.