Head Feature Extraction for Orientation-Resilient Target Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face detection technologies in passenger flow statistics struggle with accurately counting individuals when their faces are not directly facing the camera, necessitating the extraction of head features for effective people counting.
Innovation Solution
A target detection method utilizing a series of convolutional layers, including single-module and double-module residual convolution layers, to extract and enhance features from input images, allowing for accurate detection of targets such as human heads without adding additional layers, thereby reducing computational burden and enhancing real-time performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional face detection algorithms are used for passenger flow statistics, then detection accuracy is improved when faces are directly facing the camera, but detection accuracy deteriorates when faces are not directly facing the camera
Solution Approach 1:
The patent changes the detection parameter from face-specific features to head feature parameters. By extracting head features (shape, size, position) instead of relying on face orientation and expression, the system maintains high detection accuracy across various orientations and conditions, resolving the contradiction between detection accuracy and adaptability to different face orientations
Solution Approach 2:
The patent develops a detection model that serves multiple functions: it can detect heads regardless of face orientation, lighting conditions, or expression. This universal head detection approach replaces the specialized face detection algorithm, enabling the system to handle diverse scenarios (front-facing, side-facing, obscured faces) with a single unified model, thus improving both accuracy and adaptability
2Measurement precision
If additional convolution layers are added to increase receptive field, then detection accuracy is improved, but computational burden increases
Solution Approach 1:
Instead of increasing depth by adding more convolution layers, the patent expands the receptive field by increasing the spatial dimension (kernel size). Using a larger single convolution kernel (e.g., 7x7 or larger) achieves the same receptive field expansion as multiple smaller kernels stacked in depth, but with fewer parameters and lower computational complexity, thus improving detection accuracy without increasing device complexity
Solution Approach 2:
The patent uses a shallow copy approach where the head detection model reuses feature extraction capabilities from pre-trained models without requiring deep additional layers. By copying and adapting existing feature extraction mechanisms rather than building entirely new deep architectures, the system achieves effective detection with reduced computational burden
3Measurement precision
If complex feature extraction processes are used to handle various head orientations, then detection accuracy is improved, but real-time performance deteriorates
Solution Approach 1:
The patent changes the detection parameters from complex multi-view geometric transformations to simple head feature parameter extraction. By focusing on invariant head properties (overall shape, size, position) rather than orientation-specific features, the system achieves accurate detection of heads in various orientations through simple parameter comparison, maintaining both high detection accuracy and real-time performance
Solution Approach 2:
The patent extracts only the essential head features needed for detection (shape, size, position) from the input images, discarding unnecessary complex orientation and expression information. This selective extraction of critical features enables fast processing while maintaining detection accuracy, as the system only needs to identify head presence and basic properties rather than analyzing detailed facial characteristics
Data Source
AI summary
A target detection method and apparatus, a computer-readable storage medium, and an electronic device. The method comprises: acquiring an input image, and performing feature extraction on the input image by means of a first convolutional layer and a second convolutional layer to obtain a reference feature image (S110); performing feature extraction on the reference feature image by means of a third convolutional layer to obtain a first target feature image, and performing feature extraction on the first target feature image by means of a fourth convolutional layer to obtain a second target feature image (S120); acquiring a reference pre-selected image of each point on the first target feature image and the second target feature image (S130); and determining a target pre-selected image from among a plurality of reference pre-selected images, and completing target detection by means of a target detection algorithm (S140).