Object Detection Device Occlusion Ratio Threshold Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems using deep neural networks (DNNs) face reduced detection accuracy when faced with objects for which only a small number of images are available as teacher data or when objects appear differently in input images compared to the teacher data.
Innovation Solution
An object detection device that calculates confidence levels for regions in an image and determines occlusion ratios to improve detection accuracy by relaxing detection conditions for occluded objects, using a multi-layer convolutional neural network to identify occluded regions and adjust confidence thresholds accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a DNN is learned using multiple images of known objects as teacher data, then detection accuracy for known objects is improved, but detection accuracy for objects with limited teacher data or different appearances deteriorates
Solution Approach 1:
The system performs preliminary detection to identify candidate regions that may contain occluded objects before applying the occlusion handling mechanism. By pre-identifying regions with potential occlusions based on confidence scores and spatial relationships, the system prepares for subsequent occlusion ratio calculation and threshold adjustment, enabling early intervention to preserve detection accuracy for objects with limited teacher data
Solution Approach 2:
The system dynamically changes the confidence threshold parameter based on the calculated occlusion ratio. When occlusion is detected (occlusion ratio exceeds threshold), the system lowers the confidence threshold from a first threshold to a second threshold, allowing objects that would otherwise be rejected due to reduced confidence scores to be successfully detected. This parameter adaptation resolves the contradiction by maintaining detection accuracy across varying occlusion conditions
2Measurement precision
If a strict confidence threshold is applied to ensure high detection accuracy, then false positives are reduced, but occluded objects are missed
Solution Approach 1:
The system dynamically adjusts the confidence threshold based on the occlusion state of detected objects. Instead of using a fixed threshold, the threshold becomes a variable that adapts to the specific detection context. When an object is identified as occluded through the occlusion ratio calculation, the system switches from a first confidence threshold to a second, lower confidence threshold, enabling the detection system to maintain both high accuracy for clear objects and high completeness for occluded objects
Solution Approach 2:
The system implements a feedback mechanism where the occlusion ratio calculation results directly influence the confidence threshold selection. The detection process forms a closed loop: detect object → calculate occlusion ratio → adjust confidence threshold based on occlusion level → re-evaluate detection confidence. This feedback loop ensures that detection decisions are continuously optimized based on the actual occlusion state, resolving the trade-off between accuracy and completeness
Data Source
AI summary
An object detection device includes a processor configured to calculate, for each of a plurality of regions in an input image, a confidence indicating a degree of certainty that an object to be detected is represented; determine that the object is represented in a first region for which the confidence is equal to or higher than a first confidence threshold; calculate an occlusion ratio of an occluded region, in which the object is occluded in the first region by another object, to the first region; and determine that the object to be detected is represented in a second region which at least partially overlaps with the first region, when the occlusion ratio is equal to or higher than an occlusion ratio threshold and the confidence for the second region is equal to or higher than a second confidence threshold which is lower than the first confidence threshold.