Real-Time Video Target Recognition with Reused Position Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CNN structures struggle to recognize targets effectively when the target area in an image is less than 50%, particularly in applications like gesture recognition, due to high computational workload and resource intensive processes.
Innovation Solution
A method and system that utilize a cascaded neural network approach, where a first-stage neural network performs target object position detection and a second-stage neural network performs recognition, with the operation of the first-stage network optimized by reusing position detection results from previous frames to reduce CPU usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a cascaded neural network is used to perform target detection and recognition on every frame of video, then detection and recognition accuracy is improved, but CPU computational workload increases significantly
Solution Approach 1:
The patent implements periodic action by selectively executing the first-stage neural network at intervals rather than on every frame. The system determines whether to perform detection based on frame intervals and target movement characteristics, reducing CPU workload while maintaining detection accuracy for moving targets.
Solution Approach 2:
The patent applies dynamics by adaptively adjusting the detection interval based on target movement speed. When targets are detected to be moving quickly, the system reduces the interval between detections, while for stationary or slow-moving targets, the interval increases, optimizing resource usage dynamically.
2Speed
If the first-stage neural network operates at high frequency to detect moving targets, then detection speed is improved, but CPU resource consumption increases
Solution Approach 1:
The system dynamically adjusts the detection frequency based on target movement characteristics. For fast-moving targets, detection occurs more frequently to maintain tracking accuracy, while for slow-moving or stationary targets, detection frequency is reduced to conserve CPU resources.
Solution Approach 2:
The patent changes the parameter of detection interval based on target movement speed and other factors. By adjusting this parameter dynamically, the system optimizes the balance between detection speed and CPU resource consumption for different scenarios.
3Productivity
If target detection is performed on every video frame, then real-time detection capability is improved, but heat generation and hardware stress increase
Solution Approach 1:
The patent implements periodic detection rather than continuous detection on every frame. By using frame intervals and selective execution based on target characteristics, the system maintains real-time detection capability for moving targets while reducing overall CPU usage and heat generation.
Solution Approach 2:
The system uses information from previous detection results to inform future detection decisions. By leveraging target movement patterns and prediction from prior frames, the system intelligently determines when detection is necessary, reducing redundant computations and hardware stress.
4Measurement precision
If a deep neural network with many layers is used to improve recognition accuracy, then classification performance is improved, but computation workload increases sharply
Solution Approach 1:
The patent segments the recognition task into two stages: a first-stage neural network for target detection and position identification, and a second-stage neural network for detailed recognition and classification. This segmentation allows the computationally intensive deep network to be used only when necessary, reducing overall computation workload while maintaining high accuracy.
Solution Approach 2:
The first-stage neural network performs preliminary detection and identifies potential target regions before the second-stage network performs detailed recognition. This preliminary action filters out non-target regions, allowing the deep network to focus computation only on relevant areas, thereby reducing overall computation workload.
Data Source
AI summary
This disclosure provides a method and a system for detecting and recognizing a target object in a real-time video. The method includes: determining whether a target object recognition result RX-1 of a previous frame of image of a current frame of image is the same as a target object recognition result RX-2 of a previous frame of image of the previous frame of image; performing target object position detection in the current frame of image by using a first-stage neural network to obtain a position range CX of a target object in the current frame of image when the two recognition results RX-1 and RX-2 are different; or determining a position range CX of a target object in the current frame of image according to a position range CX-1 of the target object in the previous frame of image when the two recognition results RX-1 and RX-2 are the same; and performing target object recognition in the current frame of image according to the position range CX by using a second-stage neural network. Therefore, the operating frequency of the first-stage neural network used for position detection is reduced, the recognition speed is accelerated, and the usage of CPU and internal memory resources is reduced.