Scale-Equivariant CNN Object Recognition for Variable-Size Traffic Targets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current driver assistance and semi-automated driving systems face challenges in accurately tracking and recognizing objects in images, particularly in mixed traffic scenarios where objects of varying sizes and relevance require distinct processing, leading to inefficiencies in trajectory planning and resource allocation.
Innovation Solution
The method employs convolutional neural networks with filters convertible by scaling, enabling size-invariant or size-equivariant processing to decouple semantic correspondence from size, allowing for the detection of specific objects regardless of size changes, and focuses processing resources on traffic-relevant objects by interpolating object size and evaluating distance, thereby enhancing object tracking and trajectory planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional convolutional neural networks are used for object recognition, then object detection can be performed, but size variations of objects lead to inaccurate recognition and tracking
Solution Approach 1:
The patent applies parameter changes by transforming filters through scaling operations to adapt to different object sizes. The filters are converted into multiple scaled versions that can match objects at various scales in the input image, enabling size-invariant object recognition while maintaining high detection accuracy
Solution Approach 2:
The patent implements universality by creating a set of filters that can serve multiple size scales simultaneously. A single base filter is transformed into multiple scaled filters that can collectively detect objects of different sizes, making the recognition system universally applicable across various object dimensions without requiring separate detectors for each size
2Adaptability or versatility
If multiple filters convertible by scaling are used to improve size adaptability, then object recognition across different sizes improves, but memory requirements and computational complexity increase
Solution Approach 1:
The patent applies segmentation by dividing the filter set into multiple scaled versions of a base filter. Instead of using entirely different filters for each scale, the filter bank is segmented into systematic scaled variants, reducing the overall number of independent parameters while maintaining adaptability across size scales
Solution Approach 2:
The patent uses copying by generating multiple scaled versions of a single base filter through scaling operations. Rather than training completely independent filters for each scale, the system creates copies of the base filter at different scales, significantly reducing memory requirements and training complexity while preserving size adaptability
3Productivity
If all objects in the image are processed equally, then comprehensive detection is achieved, but processing efficiency decreases when focusing on traffic-relevant objects
Solution Approach 1:
The patent applies local quality by differentiating the processing approach based on object characteristics. Traffic-relevant objects receive enhanced processing attention through the size-adaptive filter matching, while less relevant objects are processed with standard methods, optimizing computational resources according to local object importance
Data Source
AI summary
A method for recognizing at least one object in at least one input image. In the method, a template image of the object is processed by a first convolutional neural network (CNN) to form at least one template feature map; the input image is processed by a second CNN to form at least one input feature map; the at least one template feature map is compared to the at least one input feature map; it is evaluated from the result of the comparison whether and possibly at which position the object is contained in the input image, the convolutional neural networks each containing multiple convolutional layers, and at least one of the convolutional layers being at least partially formed from at least two filters, which are convertible into one another by a scaling operation.