The purpose of the onboard feature indexing system is to perform pattern recognition and data compression onboard. We use the unsupervised machine learning algorithm kmeans to classify the lidar profile data and generate an index dictionary. Then we train the Radial Basis Function neural network with the index dictionary on ground computers. Finally, we use the same RBF model for the onboard feature recognition and indexing. We implemented a prototype of the onboard computer with ZISC (Zero Instruction Set Computing) chips and FPGA (Field Programmable Gate Array) so that it takes advantage of intrinsic parallel computing and reconfigurability. We tested a set of 44K profiles as the training set to learn prototypical profiles that make up the indexing dictionary. With 64 indices, we reach a high compression rate 99.17% with reasonable error range. We found the required neurons are equal to the indices. We also compared our method to wavelet algorithm and found that it significantly outperforms the wavelet compression technique.
[1]
H. Simon,et al.
EPAM-like Models of Recognition and Learning
,
1984,
Cogn. Sci..
[2]
Christopher M. Bishop,et al.
Neural networks for pattern recognition
,
1995
.
[3]
Russell Tessier,et al.
c ○ 2001 Kluwer Academic Publishers. Manufactured in The Netherlands. Reconfigurable Computing for Digital Signal Processing: A Survey ∗
,
1999
.
[4]
Robert M. Gagliardi.
On-Board Processing
,
1991
.
[5]
Loren Lemmerman,et al.
Needs for communications and onboard processing in the vision era
,
2002,
IEEE International Geoscience and Remote Sensing Symposium.