Data processing: vehicles – navigation – and relative location – Vehicle control – guidance – operation – or indication – Vehicle subsystem or accessory control
Reexamination Certificate
2000-01-04
2001-07-24
Nguyen, Tan (Department: 3661)
Data processing: vehicles, navigation, and relative location
Vehicle control, guidance, operation, or indication
Vehicle subsystem or accessory control
C701S046000, C706S025000
Reexamination Certificate
active
06266593
ABSTRACT:
TECHNICAL FIELD
The invention relates to automotive occupancy sensing (AOS) systems and methods for sensing and determining the occupancy state, including the nature, type or location of occupant (if any) with respect to the vehicle interior, and more particularly with respect to the occupant seat and dashboard or instrument panel (IP), to develop a signal useable by the airbag deployment system (ADS) by which the ADS can deploy or not (abort deployment), or modify deployment for multiphase airbags, or for partial or controlled rate inflation airbags, collectively herein termed Smart Airbag Systems. The system is characterized by defining in the vehicle interior an Occupancy Zone (OZ), a Keep-Out Zone (KOZ) and a “Gray Zone” (GZ) intermediate to the OZ and KOZ. Sensor signals from objects/persons in the Gray Zone area selectively used or discarded as part of the conditioning and training process for a neural net AOS process, and thereby detection performance is substantially enhanced.
BACKGROUND OF THE INVENTION
For background on AOS systems see Corrado et al., U.S. Pat. No. 5,482,314. Such systems produce a signal for input to the ADS, which if the occupant is out of position (OOP) or in a rear facing infant seat (RFIS) (in the front seat of a vehicle), the deployment of the airbag is aborted, deferred or otherwise controlled, as in SAS.
More recent studies have revealed that there is a class of slow speed automotive accidents causing injury to children, youngsters and frail adults. This usually occurs when the &Dgr;V of the “crash” is 18 miles per hour or less, where the occupant or RFIS is unbelted and the driver jams on the brake. The airbag deployment sensor experiences a G-force great enough to signal deployment. During these low speed accidents, the child or occupant typically has slid, or is sliding, forward into the IP and KOZ when the airbag deploys. The airbag deployment injures the child because it is too close, having intruded into the Keep Out Zone (KOZ).
Neural net algorithms have a wide range of successful useful applications in various technical fields, such as voice recognition, diagnostic systems and machine vision, and are generally well known in the art. Neural networks are sometimes used in system classification tasks for which not all the rules, system descriptions, or underlying equations are known. For such systems, the neural network is presented with a limited number of samples or “snapshots” (neural network inputs), and the known system outputs for those snapshots (neural network outputs). This input and output data is collectively known as the neural network “training set”, and is used to train (or program, or “adapt”) the neural network to the particular problem or system it is supposed to solve or identify. If the training set, the neural network architecture, and the training rules and parameters are chosen properly, a system capable of correctly identifying patterns of data typically not originally present in the training set can be built. Here lies the power of neural-network-based systems: they are able to generalize and infer correct results from a limited (original) training dataset.
The process of developing an adaptive type of neural net algorithm is typically a 3 step process: Step 1 is to generate a data set that contains the test cases the system is required to learn. Step 2 is to train the neural network to recognize the data and separate it into the required decision space. Step 3 is to is to evaluate the performance against data not included in the training data set used in step 1.
SUMMARY OF THE INVENTION
The invention is preferably employed in association with a hierarchical discrimination system as a gateway to the probability analysis disclosed in our prior U.S. Pat. No. 5,482,314, and is preferably used in association with ultrasound sensing, or other sensor type, to determine the intrusion into a defined Keep Out Zone (KOZ) between the instrument panel and the occupant seat. The invention comprises establishing a Gray Zone (herein GZ) formed by a zone intermediate the KOZ and the occupancy zone (herein OZ) which is the proper seating area of a passenger, and discarding selected data from incursions, or data from selected types or scenarios of incursions, into the GZ during neural net training (and optionally during operation). Unexpectedly, by excluding signals from the GZ, particularly and preferably when training the neural net, the performance in working practice is improved by around 5% or better, which is significant as this refers to statistical probability percentages of recognition, of the type disclosed in our prior U.S. Pat. No. 5,482,314.
AOS systems, such as that disclosed in U.S. Pat. No. 5,482,314, may make use of neural net algorithms as an element of the signal processing and occupancy state determination methodology. The disclosure of that patent is hereby incorporated by reference to the extent needed, e.g., particularly for discussion of fusion algorithms, state determinations, probabilities, and the like. The GZ can be defined by an overlap in the OZ (also known as the Non-KOZ) and the KOZ, or by separating the OZ from the KOZ.
Neural networks may be advantageously used as elements of AOS systems. The training characteristics described above are useful, since it is not generally desirable to train or configure the AOS system based on a very large number of seat occupancy scenarios. Instead, a limited set of training vectors (consisting typically of 30,000 individual seat occupancy scenarios) is used to train an AOS neural network system. The AOS neural network then uses this information to correctly classify and identify seat occupancy cases it has not seen before (unknown cases).
The generalization capability of a neural network, however, is also limited. In this lies one of the disadvantages of using neural networks as system classification tools: even though they perform well against a much larger set of test cases (unknowns in later operation) than the original training data set (from the training vector scenarios), they do not perform well for ALL test cases. This point is of particular importance to safety-critical AOS systems, since such systems must perform well during all possible occupancy scenarios. Preferably alternate hardware means of detection of these cases are employed. Such means typically include additional information input sources, such as motion detectors, or an additional occupancy sensor input.
A second disadvantage of neural networks is that once a system is “trained,” it performs well, but is typically difficult to analyze. It behaves as a “black box” (albeit a black box that does what it is trained to do), giving the user little insight on how this system arrives at its results. This is related to the architecture of neural networks, typically consisting of many interconnected neurons or processing elements, with each connection characterized by a specific weight, which by itself, does not mean much. It is only the combination of these weights and interconnections that give neural networks their power as classification and system identification tools.
The key to the use of neural networks in any classification task in general, and in AOS classification systems in particular, lies in “conditioning” the neural network, by which is meant more than applying well known Training Rules, but also selection of the training set and removal of signal ambigities in one or more zones, and/or in an intermediate GZ. Thus, the method of the invention comprises improvements in the neural net “training” protocol which involves steps of conditioning the neural net with less that all of the sensor data. More specifically the invention involves improvements in the selection of the training set and the added step of removal of ambiguous GZ signals during training. The preferred method of the invention for training neural networks for AOS classification systems comprises the following conditioning steps for a selected neural network Architecture and set of Training Rules:
1. Select a representative kno
Decker Steven W.
Seip Ralf
Dulin, Esq. Jacques M.
Innovation Law Group Ltd
Nguyen Tan
Robert Bosch Corporation
LandOfFree
Automotive occupancy sensor gray zone neural net... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Automotive occupancy sensor gray zone neural net..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Automotive occupancy sensor gray zone neural net... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2460468