Fisheye Camera Human Detection Using Height Threshold Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing human detection techniques in factory automation using deep learning are inefficient for real-time analysis due to high processing loads and inaccuracies in detecting humans amidst movable objects like corrugated cardboard pieces.
Innovation Solution
An information processing apparatus equipped with a movable object detector, a human determiner, and a human detector, which uses a fisheye camera to detect movable objects and determine if they are humans by comparing the distance between predetermined points on the object's outline with a threshold range based on human height, thereby reducing processing load and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is used for human detection, then detection accuracy is improved, but processing time increases and real-time analysis becomes difficult
Solution Approach 1:
The detection process is segmented into multiple stages: first detecting movable objects using simple motion detection, then filtering candidates based on height thresholds, and finally applying deep learning only to confirmed human candidates. This segmentation reduces the overall processing time while maintaining detection accuracy.
Solution Approach 2:
Preliminary filtering is performed before deep learning detection by detecting movable objects and checking their heights against predetermined thresholds. This preliminary action eliminates non-human objects early in the process, reducing the computational burden of subsequent deep learning operations.
2Productivity
If the target detection area is limited to movable object areas, then processing load is reduced, but detection accuracy decreases due to inclusion of non-human movable objects
Solution Approach 1:
Different quality checks are applied at different stages: motion detection for initial candidate selection, height threshold filtering for candidate refinement, and deep learning for final confirmation. Each stage applies appropriate quality criteria to the specific set of candidates at that stage, optimizing both efficiency and accuracy.
Solution Approach 2:
The detection process uses parameter changes by adjusting the stringency of detection criteria at different stages. Initial motion detection uses loose criteria to capture all potential targets, then height parameters are used to filter candidates, and finally deep learning applies strict verification only to remaining candidates, optimizing the balance between speed and accuracy.
3Device complexity
If object detection uses shape information, then processing is simplified, but detection accuracy decreases for humans who change shape with posture
Solution Approach 1:
Height measurement is performed as a preliminary action before detailed shape analysis. By measuring the vertical extent of movable objects and comparing against predetermined human height thresholds, the system filters out non-human objects early, allowing subsequent deep learning to focus only on human-shaped candidates regardless of posture variations.
Data Source
AI summary
An information processing apparatus includes a movable object detector that detects a movable object from a captured image captured with a fisheye camera, a human determiner that determines whether the movable object is a human by comparing a distance between two predetermined points on an outline of a movable object area including the movable object with a threshold range set based on a height of the human measured at a position of the movable object in the captured image, and a human detector that detects the human from the movable object area including the movable object determined as the human by the human determiner.