Human Detection in Crowds via Depth Height Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Human detection in high density crowds is challenging due to occlusions and differences between training datasets and real-world scenarios, particularly in images where one body may obscure another, and existing methods rely on learning techniques that are not effective in such complex environments.
Innovation Solution
The use of depth-sensing cameras to capture pixel depth information, convert it into height values above a ground surface, and identify head and shoulder regions through a non-learning based process, employing a multiple threshold technique and mean-shift method to detect individuals without relying on artificial intelligence or learning techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learning techniques (Histogram of Oriented Gradients, Deformable Part-based Model, Convolutional Neural Networks) are used for human detection, then recognition quality may be improved in some scenarios, but detection accuracy deteriorates in high density crowds due to occlusions and differences between training datasets and real scenarios
Solution Approach 1:
The patent segments the human detection problem into two distinct parts: head detection and shoulders detection. By separately identifying head regions and shoulder regions and then determining their spatial relationship, the system avoids the pitfalls of trying to detect full-body human instances in crowded scenes. This segmentation allows each detection task to be optimized independently, improving overall reliability in high density crowds where occlusions are common.
Solution Approach 2:
The patent transitions from traditional 2D image space analysis to 3D spatial reasoning by incorporating depth information from depth-sensor cameras. By converting depth values to height values above ground surface and using spatial relationships between head and shoulder regions in three-dimensional space, the system achieves more reliable detection in scenarios where 2D methods fail due to occlusions and viewpoint variations.
2Loss of information
If full-body based human detection is used, then comprehensive human information can be obtained, but detection becomes difficult in high density crowds where one body may occlude another
Solution Approach 1:
The patent divides the human body into separate detectable parts (head and shoulders) that can be identified independently even when the full body is occluded. By detecting head regions and shoulder regions separately and establishing their spatial relationship, the system recovers human information that would be lost in full-body detection approaches, particularly in high density crowds where bodies overlap and occlude each other.
Solution Approach 2:
The patent uses depth information from depth-sensor cameras as an intermediary to facilitate detection in crowded scenes. The depth data provides additional spatial cues that help distinguish overlapping bodies and determine the spatial relationship between head and shoulder regions, enabling detection even when visual occlusion prevents traditional full-body detection methods from working.
3Measurement precision
If depth-sensor cameras are used to capture pixel depth information and convert to height values, then detection accuracy in crowds is improved, but device complexity increases
Solution Approach 1:
The patent employs depth-sensor cameras that serve multiple functions: capturing both standard color image data and depth information simultaneously. This multi-functionality allows the system to leverage depth data for improved crowd detection accuracy while using the same hardware infrastructure for other surveillance tasks, thereby mitigating the increase in device complexity through efficient resource utilization.
Data Source
AI summary
The present disclosure describes a non-learning based process and apparatus for detecting humans in an image. This may include receiving an image that has pixel distance information from a camera and using that to determine a height of the pixel above a ground surface. One or more regions may then be identified that may include a head and shoulders of an individual in the image. A multiple threshold technique may be used to remove some background regions, and a mean-shift technique used to find the local highest regions that may be combination of head and shoulders of the person. In embodiments, the view angle and/or the height of the camera may not be fixed.