Image Processing with Neural Density Maps for Crowd Flow Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for measuring the flow of target objects in crowded scenes, such as crowds of people, suffer from reduced accuracy due to overlapping individuals, leading to inaccuracies in detecting and tracking individuals.
Innovation Solution
An image processing apparatus utilizing a neural network-based approach to estimate density distributions and positional coordinates of target objects, followed by tracking and measuring their movement loci to improve accuracy in crowded scenes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional detection methods are used to detect target objects in crowded scenes, then the detection process is simple, but the measurement precision deteriorates due to overlapping individuals
Solution Approach 1:
The patent introduces density distribution maps as an intermediary representation between the input image and target object detection. Instead of directly detecting overlapping individuals, the system first generates a density map that distributes target presence information across the image space, effectively mediating the detection process and resolving overlapping issues through spatial distribution analysis
Solution Approach 2:
The patent segments the detection process into multiple independent components: density map generation, peak detection, and coordinate extraction. By dividing the complex task of detecting overlapping individuals into these sequential segments, each component can be optimized independently, improving overall detection precision without proportionally increasing system complexity
2Measurement precision
If visual tracking models are used to track individuals in crowd scenes, then tracking capability is provided, but measurement precision deteriorates due to difficulty in distinguishing overlapping individuals
Solution Approach 1:
The patent applies partial action by focusing tracking efforts only on detected peak positions in the density map rather than attempting to track all potential individual positions. This selective approach concentrates computational resources on high-probability target locations, improving tracking accuracy while reducing the difficulty of distinguishing overlapping individuals
Solution Approach 2:
The patent performs preliminary density map generation and peak detection before initiating the tracking process. By pre-identifying target positions through density analysis, the system establishes accurate initial tracking points that prevent the common error of misidentifying overlapping individuals, thereby improving subsequent tracking precision
3Measurement precision
If density distribution estimation is used to handle overlapping individuals, then measurement precision of target positions is improved, but device complexity increases due to additional processing steps
Solution Approach 1:
The patent replaces traditional mechanical detection approaches with a field-based density distribution model. Instead of directly detecting individual targets through mechanical or geometric methods, the system substitutes a continuous density field representation that naturally handles overlapping cases through mathematical distribution, improving position estimation accuracy while the modular implementation keeps processing complexity manageable
Solution Approach 2:
The patent changes the fundamental parameter representation from discrete individual detections to continuous density distribution values. By transforming the detection problem into a density estimation problem with different parameter characteristics, the system achieves improved position estimation accuracy for overlapping individuals while the parameter transformation itself provides a systematic processing framework
Data Source
Figure 1~2
Figure 3
Figure 4A~4B
AI summary
There is provided with an image processing apparatus (100). An obtaining means (201) obtains a moving image including a plurality of images. A density estimating means (202) estimates a density distribution of a target object based on an image of the plurality of images. A position estimating means (203) estimates a position of the target object in the image based on the density distribution. A measuring means (205) measures a flow of the target object from positions of the target object in the plurality of images.