1 / 16

Distributed and Efficient Classifiers for Wireless Audio-Sensor Networks

Baljeet Malhotra Ioanis Nikolaidis Mario A. Nascimento University of Alberta Canada. Distributed and Efficient Classifiers for Wireless Audio-Sensor Networks. Supported By:. To be presented at: INSS, June 17-19, 2008, Kanazawa, Japan. Outline Of The Talk. Introduction Motivation

lacey-kidd
Télécharger la présentation

Distributed and Efficient Classifiers for Wireless Audio-Sensor Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Baljeet Malhotra Ioanis Nikolaidis Mario A. Nascimento University of Alberta Canada Distributed and Efficient Classifiers for Wireless Audio-Sensor Networks Supported By: To be presented at: INSS, June 17-19, 2008, Kanazawa, Japan

  2. Outline Of The Talk • Introduction • Motivation • Classification of Acoustic Targets • Classification Framework • Classification Methods: KNN & ML • Features Extraction • Independent Features Selection • Global Features Selection • Simulation Study • Dataset and Setup • Methodology • Results and Discussions • Conclusion and Future Directions 2 min 5 min 5 min 6 min 2 min

  3. Introduction • Vehicle classification is an important problem in WSN • Tracking • Localization • Tracking can be facilitated by: • Video/Image based sensors • RFID tags • Limitations: • Video/Image requires higher processing capabilities • RFID tags may not be feasible • Acoustic target tracking • Lesser processing requirements

  4. Vehicle Classification • Vehicle classification is crucial to tracking • Only vehicles of interest are reported • Problem becomes much challenging if there are more vehicles of the same class • Identification problem • This paper deals with the problem of vehicle classification only and NOT identification Class A Class B Class A Disclaimer: Images used above are collected through Google’s search engine

  5. A Framework for Classification • Nodes organize themselves into neighborhoods “clusters” as a vehicles crosses through an area monitored by sensors • A master node is selected based on the signal strength. • A cluster can perform classification independently. • Multiple clusters may be formed and collaborate for: • Better accuracy • Sharing the costs • But not attempted in this paper (future work) Formation of a cluster Sensor deployment along a straight path

  6. Classification Techniques • k-NN is one of the simplest, yet accurate methods. • Given a set of samples known samples, U • Fetch k (≥ 1) closest known samples from U • Classifies the unknown sample as the majority class of the drawn k samples. • Maximum Likelihood (ML) • Real time computation is proportional to: • d × l × c (for KNN) • d2 (for ML) • d : size of feature vectors, l : class size, c : number of classes • Conclusion: Features vector size is important

  7. Feature Extraction • Hundreds of features to choose from acoustic signatures • Two demands that compete with each other • Low dimensional features that are yet effective • Acoustic features • Power spectral density • Power is concentrated in the lower range of frequencies Dragon Wagon Assault Amphibian Vehicles

  8. Feature Extraction Schemes • Pruning Step 1: Select the frequencies that have the maximum power as reported by training samples: • where • Pruning Step 2: Ranking and selecting only a % of them: • (< ) • Independent Feature Selection • a • Global Feature Selection • s

  9. Experimental Study • DAPRA/IXO SenseIT dataset • Two types of vehicles (AAV and DW) • Total 389 samples (180 AAV, and 209 DW) • Simulated a network of (3 ~ 40) sensors • In order to create a local copy of unknown (testing) sample for a sensor, a signal is attenuated based on its distance from the moving vehicle, and white noise is added • Performance Metrics • Classification accuracy • Communication (energy) expenditure

  10. Evaluation Methodology • Classification accuracy: • Based on leave-one-out policy • Energy expenditure model: • Er = 50nJ/bit and Es = 50+.1×R3 nJ/bit, where Er is the energy required to receive one bit and Es is the energy required to send one bit at R distance. • L1 Distance Metric

  11. Evaluation of Results Size of IFS and GFS Feature Vectors IFS GFS • Size does not go beyond 20 and 15 in IFS and GFS respectively

  12. Evaluation of Results Classification Accuracy KNN ML • ML outperforms KNN

  13. Evaluation of Results Communication Costs KNN ML • DEF is less expensive than DAF

  14. Evaluation of Results Comparison with other studies

  15. Conclusion and Future Direction • Classifying ground vehicles is an important problem in wireless sensor networks. • We have two main contributions in this work: • Distributed data/decision fusion framework for classification • New feature extraction schemes that can produce low dimensional yet effective features • We conducted a simulation study using real acoustic signals of military vehicles, and our proposed features achieved better classification accuracy • In the future: • Improve the efficiency of our proposed schemes. • Consider more than two classes of ground vehicles

  16. Thank You !

More Related