Video Object Detection Using Keyframe Preselection Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional target detection methods in video processing are inefficient for high-resolution videos above 720p due to excessive anchor points, leading to low detection speed and reduced accuracy, which can only be mitigated by frame skipping at the cost of detection accuracy.
Innovation Solution
The method involves dividing video frames into keyframe and non-keyframe images, recognizing objects in keyframes, scaling detection boxes for non-keyframes, cropping non-keyframes to generate preselection boxes, and processing these boxes through a non-keyframe detection network to optimize detection speed without compromising accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If traditional target detection method processes each video frame with full anchor point classification, then detection accuracy is maintained, but detection speed becomes too slow for real-time processing
Solution Approach 1:
The patent segments video frames into keyframes (processed with full detection) and non-keyframes (processed with accelerated detection using preselection boxes). This segmentation allows the system to apply different processing strategies to different frames, reducing overall time consumption while maintaining detection accuracy across the video sequence.
Solution Approach 2:
The patent performs preliminary detection on keyframes to generate detection boxes, which are then scaled and used as preselection boxes for non-keyframes. This preliminary action on keyframes provides prior information that accelerates processing of subsequent non-keyframes without requiring full anchor point classification, thus reducing time consumption.
2Productivity
If frame skipping is used to increase detection speed, then processing time is reduced, but detection accuracy deteriorates
Solution Approach 1:
The patent implements periodic full detection on keyframes (e.g., every Nth frame) while using accelerated detection on intermediate non-keyframes. This periodic application of full detection ensures detection accuracy is maintained at regular intervals, while the accelerated processing on non-keyframes increases overall throughput without significant accuracy loss.
Solution Approach 2:
The detection results from keyframes provide feedback that informs the processing of non-keyframes. The detection boxes from keyframes are scaled and used as preselection boxes for non-keyframes, creating a feedback loop that maintains accuracy while enabling accelerated processing through intelligent reuse of prior detection information.
3Reliability
If full anchor point classification is performed on all frames, then detection accuracy is maintained, but computational complexity increases
Solution Approach 1:
The patent applies different quality levels of detection to different frames: full anchor point classification (high quality) is applied to keyframes, while accelerated detection with preselection boxes (reduced quality but sufficient for many cases) is applied to non-keyframes. This local differentiation of quality levels maintains overall detection reliability while reducing average computational complexity.
Solution Approach 2:
For non-keyframes, the patent performs partial detection action by using scaled preselection boxes from keyframes instead of full anchor point classification. This partial action is sufficient for maintaining detection reliability in most cases while significantly reducing computational complexity compared to exhaustive processing of every frame.
Data Source
AI summary
A method for accelerated detection of objects in videos, a server, and a non-transitory computer readable storage medium are provided. The method realizes the detection of a target object in a video by dividing all frame images in video images into preset groups of frame images, each group of frame images including a keyframe image and a non-keyframe image, using a detection box of a target in the keyframe image to generate a preselection box in the non-keyframe image, and detecting the location of the target in the preselection box.