Camera Image Conversion for Intrusion Detection Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing monitoring systems for detecting intrusions into no-entry zones, such as railroad platforms, face accuracy issues due to camera shooting region deviations, requiring parameter re-adjustment or re-training of image recognition models, which is cumbersome and inefficient.
Innovation Solution
A monitoring device and system that performs an image conversion operation to align the current camera image with the original shooting region, allowing accurate vehicle detection using an image recognition model without re-adjusting parameters or re-training, even when the camera shooting region deviates within acceptable levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an image recognition model is used for vehicle detection, then detection accuracy is improved, but when camera shooting region deviates, the model requires parameter re-adjustment or re-training which increases operational complexity
Solution Approach 1:
The patent introduces an image conversion operation as an intermediary step between camera image capture and model inference. This conversion operation transforms the current shooting region image back to the original shooting region image, allowing the pre-trained model to process images as if the camera were properly positioned, thereby eliminating the need for parameter re-adjustment or re-training when minor deviations occur
Solution Approach 2:
The patent performs image conversion operations in advance before the model inference step. By pre-converting the current shooting region image to match the original shooting region, the system prepares the input data in the format the model expects, avoiding the need for post-deployment model adjustments when camera deviations occur
2Adaptability or versatility
If camera shooting region deviates from original region, then field of view changes, but re-training the model with collected data over time is troublesome and reduces productivity
Solution Approach 1:
The patent creates a converted copy of the current shooting region image that replicates what the original shooting region image would look like. This converted image serves as a substitute for the actual original image, allowing the model to process images with the same characteristics as during training, thereby maintaining high detection accuracy without requiring actual re-training
3Measurement precision
If parameters of image recognition model are re-adjusted to recover accuracy, then detection accuracy is restored, but the process is troublesome and time-consuming
Solution Approach 1:
The patent replaces the mechanical process of model parameter re-adjustment with an automated image conversion operation. Instead of manually or computationally adjusting model parameters to adapt to camera deviations, the system automatically transforms the input image to compensate for the deviation, achieving the same goal of maintaining accuracy without the time-consuming parameter adjustment process
Data Source
AI summary
Provided is a monitoring system configured such that, even when there is a deviation of a shooting region of a camera, if the deviation is within an acceptable level, an accurate vehicle detection using a recognition model can be performed without readjusting parameters or re-training the model. Based on an image of a no-entry zone and a platform zone, a processor performs a first operation for detecting a person in the no-entry zone and a second operation for detecting a vehicle in the no-entry zone; and based on detection results, determines whether an alert needs to be issued. When there is a deviation of the shooting region of a camera, the processor performs a conversion operation for converting an image of a current shooting region to an image that would be captured by the camera with an original shooting region and uses the converted image for the two detection operations.