Crowd Counting via Tiled CNN Head Prediction Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting and counting individuals in crowds, particularly those with high density, tend to underestimate the number of individuals due to issues such as small box elimination, superposition errors, and merging of close boxes.
Innovation Solution
A method implemented by computer that uses a convolutional neural network previously trained on images of crowds with annotated heads, where the annotations are modified using a tiling process as adjacent cells, to generate prediction maps and binarize them for accurate head detection and counting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional box-based detection methods are used for crowd counting, then the detection process is simple and fast, but the counting accuracy deteriorates in dense crowds due to box merging and superposition errors
Solution Approach 1:
The patent segments the crowd detection problem into two independent tasks: density map estimation (continuous crowd distribution) and instance detection (individual head localization). This segmentation allows each task to be optimized separately, with the density map providing accurate crowd-level statistics and instance detection providing individual counts, thereby resolving the accuracy-complexity contradiction.
Solution Approach 2:
The patent introduces a density map dimension as an intermediate representation between the input image and the final count. This density map serves as a continuous approximation of crowd distribution, adding a new dimensional layer that captures crowd density information without relying solely on discrete bounding boxes, thus improving accuracy while managing complexity.
2Reliability
If the detection boxes are made larger to ensure individual detection, then detection robustness improves, but box merging increases and counting accuracy deteriorates
Solution Approach 1:
The patent combines two complementary detection approaches: density map estimation (which is robust to box size variations and provides accurate crowd-level statistics) and instance detection with non-maximum suppression (which provides individual head localization). By merging these two methods, the system achieves both detection robustness and counting accuracy, as each method compensates for the other's weaknesses.
Solution Approach 2:
The patent employs non-maximum suppression (NMS) as a feedback mechanism that iteratively refines the detection results. The NMS process uses the confidence scores and overlap information from initial detections to suppress redundant boxes, providing feedback that eliminates false positives and merging errors, thereby maintaining both robustness and accuracy.
3Measurement precision
If multiple detection algorithms are combined to improve accuracy, then counting precision improves, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the computational workload into two parallel streams: density map estimation (which operates on the entire image and provides crowd-level statistics) and instance detection (which focuses on localized head regions). This segmentation allows efficient resource utilization, as each stream can be processed independently and in parallel, maintaining high accuracy while optimizing processing speed.
Solution Approach 2:
The patent applies partial action by using the density map to guide instance detection. Instead of performing full instance detection on the entire image, the system uses the density map to identify high-density regions where instance detection is most needed, thereby reducing unnecessary computations in low-density areas and improving overall processing efficiency.
Data Source
AI summary
A method, implemented by computer, for locating and counting individuals in a crowd, said method takes, as input data, one or more images of a crowd of individuals, and provides, as output data, one or more binary images of connected components corresponding to the heads of the individuals of the crowd including providing a convolutional neural network previously trained on a training set of images of crowds of individuals whose heads are annotated, the annotations having previously been modified using a process of tiling as adjacent cells, generating, for each image of a crowd, a prediction map processing said image through the convolutional neural network, binarizing each prediction map using a binarization module configured to generate a threshold value or a map of threshold values specific to said prediction map PM, each binarized prediction map being a binary image of connected components corresponding to the heads of the individuals.