Dynamic Confidence Threshold Adjustment for Object Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Object detection systems often misidentify the type of an object due to common features across different types, leading to failure in tracking and loss of detected objects over time-series images.

Innovation Solution

An object detection device that calculates confidence scores for each type of object in a series of images, tracks detected objects, updates confidence thresholds based on frequency of occurrence, and adjusts classes for non-maximum suppression processing to improve type estimation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a fixed confidence threshold is used for object detection, then the detection process is simple and fast, but the estimation accuracy of object type deteriorates when objects have common features

Engineering Contradiction:

Improveobject type estimation accuracyVSAvoidconfidence threshold adjustment mechanism

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The confidence threshold is transformed from a fixed value to a dynamic parameter that automatically adjusts based on the frequency of occurrence of different object types during tracking. The system monitors which object types are detected most frequently and lowers the threshold for those types, allowing the detection system to adapt to the specific context and improve estimation accuracy without manual intervention.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system implements a feedback mechanism where the detection results are used to update the confidence threshold. By counting the frequency of occurrence for each object type during the tracking period and using this information to adjust the threshold, the system creates a closed-loop control that continuously improves its performance based on actual detection accuracy.

Inventive Principle:

Principle #23Feedback

2Productivity

If the confidence threshold is lowered for all object types, then more objects are detected, but the accuracy of type estimation deteriorates due to false recognition

Engineering Contradiction:

Improvenumber of detected objectsVSAvoidobject type estimation accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

Instead of applying a uniform confidence threshold to all object types, the system applies different thresholds tailored to each object type based on its frequency of occurrence. Object types that appear frequently during tracking receive a lower threshold, while less frequent types maintain higher thresholds, allowing the system to optimize detection sensitivity for each category individually.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The confidence threshold parameter is changed dynamically based on the statistical characteristics of object detection. By modifying the threshold parameter according to the frequency of occurrence of each object type, the system adapts the detection sensitivity to match the actual distribution of objects in the scene, improving both detection rate and accuracy.

Inventive Principle:

Principle #35Parameter changes

3Reliability

If object detection is performed on every image in time-series, then complete tracking is achieved, but computational time and energy consumption increase

Engineering Contradiction:

Improvetracking continuityVSAvoidcomputational energy consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The system performs preliminary actions by updating the confidence threshold based on frequency of occurrence during the tracking period. This preliminary adjustment of parameters allows the system to optimize subsequent detection operations, reducing the computational burden while maintaining tracking reliability through adaptive rather than static processing.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11170272B2Object detection device, object detection method, and computer program for object detection

Publication Date: 2021.11.09 TOYOTA JIDOSHA KK

AI summary

An object detection device calculates, for each of images acquired in time-series, a confidence score of each of a plurality of types of an object to be detected for each of a plurality of regions on the image, detects the type of the object corresponding to the confidence score in a region in which the confidence score is equal to or higher than a confidence threshold of the type for each of the images, tracks the detected object, counts a frequency of occurrence for each type of the detected object in a period in which the detected object is tracked, and updates the confidence threshold in such a way that the confidence threshold of a type having a higher frequency of occurrence is lower than the confidence threshold of a type having a lower frequency of occurrence.