Vehicle Detection Using Edge Clustering and SVM Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle detection methods for forward collision warning systems are computationally intensive, making real-time detection at full VGA resolution challenging on embedded processors, and often result in lower reliability or increased cost and complexity when using additional cores or processors.
Innovation Solution
A method employing a dual-path approach using Sobel edge detection and run-length encoding (RLE) analysis, combined with summed area table (SAT) analysis, which allows for real-time vehicle detection at full VGA resolution on a single dual-core DSP by selectively using symmetry, corner, and shadow detection features, and implementing a 'bottom-up' method for detecting vertical clusters of horizontal edges to reduce processor cycles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If histogram of gradients and SVM classifier are used for object detection, then detection reliability is improved, but computational complexity increases
Solution Approach 1:
The patent segments the detection process into multiple stages: candidate generation using simple features (edges, corners, symmetry), candidate filtering using additional constraints (shadow detection, aspect ratio), and final classification using SVM only on reduced candidate sets. This segmentation reduces the number of computationally intensive operations while maintaining detection reliability.
Solution Approach 2:
The patent applies partial action by using simpler detection methods (edge detection, corner detection, symmetry analysis) for the majority of processing, reserving the computationally intensive SVM classification only for a small subset of high-probability candidates. This partial application of complex methods reduces overall computational load while maintaining reliability.
2Productivity
If additional cores or processors are used to improve detection speed, then productivity is improved, but device complexity and cost increase
Solution Approach 1:
The patent implements dynamic processing where the level of analysis applied to each candidate object varies based on its characteristics. High-probability candidates receive full SVM analysis while low-probability candidates are quickly filtered using simpler metrics. This dynamic approach optimizes processing speed without requiring additional hardware cores.
Solution Approach 2:
The patent performs preliminary filtering using computationally simple operations (edge detection, symmetry analysis, aspect ratio checking) before applying the computationally intensive SVM classifier. This preliminary action reduces the workload on main processors and eliminates the need for additional processing cores.
3Measurement precision
If full VGA resolution processing is performed, then measurement precision is improved, but computational complexity increases
Solution Approach 1:
The patent segments the image processing into hierarchical levels: first identifying candidate regions using downsampled or feature-extracted data, then applying full-resolution analysis only to promising candidates. This segmentation maintains detection precision while reducing overall computational load.
Solution Approach 2:
The patent applies different processing qualities to different regions of the image based on their likelihood of containing objects of interest. High-resolution SVM classification is applied locally only to candidate regions identified by simpler methods, rather than processing the entire VGA image at full resolution.
Data Source
AI summary
In an example detection system for vehicles or other objects of interest, objects are detected in real-time at full VGA 30 frame per second resolution. A preprocessor may perform run-length encoding (RLE) to provide detected edges. The image may then be scanned from the bottom up to identify vertical clusters or “stacks” or RLEs. Vertical clusters with low vertical density may be eliminated as poor vehicle candidates. Vertical clusters not eliminated may then be processed with a histogram of gradients algorithm, and confirmed with a support vector machine algorithm. A range to the nearest object may also be calculated, and a warning provided if the object is too close.