Range Profile Classification Using Hidden Markov Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for classifying objects using high-resolution range profiles face challenges in obtaining extensive training data and accurately comparing feature locations and amplitudes, especially when some features are not observed or additional features are present.
Innovation Solution
A method that generates probabilistic models using Hidden Markov Models (HMMs) to represent possible sequences of distances between features, allowing for classification based on the probability of observed sequences in range profiles, which accounts for uncertainty in feature observation and orientation, and includes the generation of primary and secondary probabilistic models from structural details other than direct sensor measurements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If statistical classifiers are used with extensive training measurements from all viewpoints, then classification accuracy is improved, but the complexity and time required to obtain training data increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-defining a set of characteristic features and their expected locations on object templates before actual classification occurs. These features represent physically identifiable structural elements that are likely to produce distinct peaks in range profiles. By preparing these feature locations in advance on templates, the system eliminates the need for extensive training data collection from multiple viewpoints, while still achieving accurate classification through comparison of observed peak locations with pre-defined feature locations.
2Speed
If feature locations are extracted and compared directly with model predictions, then classification speed is improved, but accuracy decreases when some features are not observed or additional features are present
Solution Approach 1:
The patent applies parameter changes by transforming the classification approach from direct feature location matching to a probabilistic framework. Instead of requiring exact matches between observed and predicted feature locations, the system uses probability distributions to represent the likelihood of feature presence and location. This allows the classifier to handle cases where some features are missing or additional features are present, maintaining both speed and accuracy by evaluating the probability of observed features given the model predictions.
3Adaptability or versatility
If extensive training data from all viewpoints is collected, then the classifier can handle all orientations, but the quantity of data required increases significantly
Solution Approach 1:
The patent applies copying by using object templates that represent the structural features of objects of interest. Instead of collecting extensive training data from multiple viewpoints, the system creates template copies with pre-defined feature locations that represent the expected positions of characteristic features. These templates serve as reference models that can be compared against observed range profiles, enabling the system to handle all orientations without requiring extensive training data for each viewpoint.
Data Source
Figure 1a~1b
Figure 2a~2c
AI summary
A method and apparatus are provided for classifying range profiles, generated for example by a radar, lidar or sonar. In the method,each in a set of objects of interest is modelled with a probabilistic model. The probabilistic model represents the probabilities of occurrence of different possible sequences of distances between selected features of the object, in different orientations, that are likely to result in peaks of backscatter in a range profile of the object. The probabilistic model is derived from a first probabilistic representation of each selected feature, generated to represent the uncertainty in locating the feature and the uncertainty in observing the feature in a range profile. Classification is achieved by calculating, for each probabilistic model, the probability that the model would generate a given sequence of distances between observed backscatter events in a given range profile. The model generating the given sequence with the greatest probability identifies the object likely to have produced the given range profile. Preferably, the probabilistic models comprise Hidden Markov Models (HMMs).