Recurrent Neural Network Object Recognition via Temporal Signatures
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition systems face computational demands due to large feature sets and lack rotational invariance, making them inefficient for identifying objects at varying orientations, especially in automotive applications.
Innovation Solution
A multi-dimensional lidar generates temporal sequences of 3D data, which are fed into a trainable recurrent neural network trained for rotational invariance by creating multiple copies of the network to process data at different angular segments, enabling recognition regardless of object orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing systems use microprocessors to process large feature sets from camera images, then object recognition can be performed, but the computational demand becomes excessively high
Solution Approach 1:
The patent extracts only the essential temporal signature features from the full image data, rather than processing all pixel information. The dynamic system selectively captures temporal variations at key locations, reducing the feature set from millions of pixels to a manageable number of temporal signatures while maintaining recognition accuracy
Solution Approach 2:
The patent segments the object recognition problem into distinct temporal phases by identifying key events in the temporal sequence. The dynamic system divides the continuous temporal data into discrete phases representing different object states or actions, making the processing more efficient and manageable
2Measurement precision
If neural networks process large numbers of features from single images, then object identification is possible, but the computational complexity increases significantly
Solution Approach 1:
The patent extracts only the essential temporal signature features from the full image data, rather than processing all pixel information. The dynamic system selectively captures temporal variations at key locations, reducing the feature set from millions of pixels to a manageable number of temporal signatures while maintaining recognition accuracy
Solution Approach 2:
The patent transitions from static image processing to dynamic temporal sequence processing. Instead of analyzing a single frozen moment, the system processes the evolving temporal patterns of features, allowing the neural network to learn from the dynamics of object appearance and behavior over time
3Device complexity
If object recognition systems process single static images, then processing is simpler, but rotational invariant recognition cannot be performed reliably
Solution Approach 1:
The patent transitions from static image processing to dynamic temporal sequence processing. Instead of analyzing a single frozen moment, the system processes the evolving temporal patterns of features, allowing the neural network to learn from the dynamics of object appearance and behavior over time
Solution Approach 2:
The patent performs preliminary training of the neural network with rotated versions of objects during the learning phase. The system pre-exposes the network to various rotational configurations, enabling it to automatically develop rotational invariance without requiring complex runtime rotation handling
Data Source
AI summary
A system for object recognition in which a multi-dimensional scanner generates a temporal sequence of multi-dimensional output data of a scanned object. That data is then coupled as an input signal to a trainable dynamic system. The system exemplified by a general-purpose recurrent neural network is previously trained to generate an output signal representative of the class of the object in response to a temporal sequence of multi-dimensional data.