Deformable Gaussian Kernels for Occlusion-Aware Crowd Counting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing crowd counting systems using traditional Gaussian kernels struggle to accurately count overlapping human heads due to occlusion, leading to inconsistent visual centers of gravity and poor training effects, resulting in inaccurate crowd density maps and significant counting errors.
Innovation Solution
Implement a method that generates training data using deformable Gaussian kernels by identifying overlapping kernels, stretching, rotating, and adjusting the center point coordinates to align the visual centers of gravity with actual picture features, ensuring feature similarity and improving the training effect of convolutional neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a traditional two-dimensional Gaussian kernel is used to generate training data, then the calculation is simple and the generation process is easy, but the visual center of gravity does not align with the actual picture features when heads are occluded, leading to poor training effects
Solution Approach 1:
The patent applies dynamics by making the Gaussian kernel deformable rather than static. The kernel dynamically adjusts its shape parameters (σx, σy) and center coordinates based on occlusion detection, allowing it to adapt to different occlusion scenarios and align with actual picture features.
Solution Approach 2:
The patent changes the parameters of the Gaussian kernel (variance σx, σy and center coordinates) based on occlusion detection. When occlusion is detected, the kernel parameters are adjusted to reflect the actual visual center of gravity in the picture, improving alignment accuracy.
2Productivity
If a fixed circular Gaussian kernel is used for all heads, then the model is simple and fast to compute, but it cannot accurately represent occluded heads where the visual center shifts due to occlusion
Solution Approach 1:
The patent applies local quality by treating different heads differently based on their occlusion status. Instead of using a uniform approach for all heads, the system detects occlusion locally and adjusts the Gaussian kernel parameters specifically for occluded heads, improving representation accuracy where needed while maintaining simplicity for non-occluded cases.
Solution Approach 2:
The system dynamically determines whether to use a fixed or deformed Gaussian kernel based on occlusion detection. This dynamic approach allows the system to maintain computational efficiency for simple cases while improving accuracy when occlusion is present.
3Ease of manufacture
If traditional Gaussian kernels are used without deformation, then the training data generation is straightforward, but the feature similarity between training data and actual pictures is low, resulting in poor training effects
Solution Approach 1:
The patent applies preliminary action by detecting occlusion and deforming Gaussian kernels during the training data generation phase rather than during actual inference. This preliminary deformation ensures that the training data already contains accurate feature representations, improving feature similarity before the model even sees the actual pictures.
Data Source
AI summary
Disclosed is a method for generating training data based on a deformable Gaussian kernel in a crowd counting system, which includes steps of: finding a set of overlapping Gaussian kernels from training data; stretching an occluded Gaussian kernel; rotating the occluded Gaussian kernel; adjusting a center point coordinate of the occluded Gaussian kernel; and determining whether there is any Gaussian kernel that have not been selected yet in the training data, and outputting a resulted crowd density map having gray values as training data. Therefore, feature similarity between the crowd density map in the training data and an actual picture is effectively increased so that a regular pattern between the training data and the actual picture can be more readily learned by a convolutional neural network and an accuracy of the crowd counting system is improved.