Predicting Human Count through Environmental Sensing in Closed Indoor Settings
Download as PPTX, PDF•0 likes•68 views
This document discusses a proposed methodology for human count detection in closed indoor environments using environmental sensing data, mitigating vulnerabilities of existing methods. The system demonstrates high accuracy (>99%) and can be discreetly deployed to prevent disruption by intruders. Future work includes identifying additional environmental factors influencing human presence and exploring outdoor applications.
Predicting Human Count through Environmental Sensing in Closed Indoor Settings
- 1. Predicting Human Count through Environmental Sensing in Closed Indoor Settings 1,3,5Department of Computer Science and Engineering 2Department of Electrical and Electronic Engineering 4Department of Civil Engineering Bangladesh University of Engineering and Technology, Dhaka, Bangladesh MobiQuitous 2018 New York City, United States Shamir Ahmed1, Uday Kamal2, Tarik Reza Toha3, Nafisa Islam4, and A. B. M. Alim Al Islam5
- 2. Outline of This Presentation • Background and motivation • Our proposed solution – Proposed methodology – Experimentation – Result analysis • Conclusion and future work 2
- 3. Background: Necessity of Human Count Detection 3 • Human count detection in a highly-secured closed indoor environment is crucial Bank vault Data centers Armory
- 4. Existing Human Counting Techniques • Different alternatives – Image recognition – Infrared detectors • A common need, which in turn engenders a common limitations – Their visibility – Prone to intentional damage by the intruders 4
- 5. • Human Tracking System Based on PIR Sensor Network and Video – Xiong et al., CWSN, 2014 – Yun et al., Sensors, 2014 • Human Sensing Using Visible Light Communication – Li et al., MobiCom, 2015 – Presents a new system named as LiSense that can produce real-time human skeleton reconstruction using Visible Light Communication (VLC) • An information technology enabled sustainability test-bed (ITEST) for occupancy detection through an environmental sensing network – Dong et al., Energy and Buildings, 2010 – Deploys a large-scale sensor network test-bed for occupancy detection in an open-plan office building – 73% occupancy using HMM Risk of being deactivated by an intruder due to its visibility Not for closed indoor settings and has low accuracy Existing Research Studies 5
- 6. Our Contribution in This Paper We propose a secured and portable system for human counting in a closed indoor settings through environmental sensing 6
- 7. Operational Block Diagram 7 Device design, development, and deployment Collect sensors data (CO, CO2, NO2, SO2, O3, LPG, Temp, and Hum) Extract environmental factors that are highly correlated with # of humans Correlation Matrix Predict human count using the classifiers’ Apply machine learning classifiers
- 8. Device Development and Deployment 8 Side view Top view Room Size (ft2) Height (ft) Inside Objects Room-1 144 13 1 table, 1 fan, 2 desktops, and 1 almirah Room-2 480 13 1 table, 2 fans, 2 desktops, and 2 almirahs Room-3 600 13 10 tables, 6 fans, 2 air conditioners, 10 desktops, 1 black board, and 3 almirahs Different sensors
- 9. Data Collection Using Our System EcoRadar 9
- 10. Correlation Matrix over Collected Data 10 # of persons present Room-1 Room-2 Room-3 CO 0.055048157 -0.005210679 -0.0158348136 CO2 0.729852349 0.834239431 0.397657642 LPG 0.794281011 0.554863591 0.482805565 NO2 0.502596592 0.0617479 0.539176588 SO2 0.477913214 0.150283339 0.574382104 O3 0.390925688 1.12436E-14 0.019222169 Temperature 0.523004245 -0.171163302 0.477701747 Humidity 0.47247865824 0.540004312 0.049569851 Sensor height 0.0086828568 0.005152112 0.0062244213 High correlation co-efficient except for CO, 03, and height
- 11. Regression Matrix over Collected Data 11 # of persons present Room-1 Room-2 Room-3 CO 0.009095953 0.935877271 0.858174808 CO2 0.00762742 0.009574814 0.144277498 LPG 9.23027E-50 3.1662E-157 0.001688016 NO2 0.161329277 0.000361105 0.176488419 SO2 7.05075E-08 0.030848214 0.001345993 O3 0.240391752 0.898772189 0.554918102 Temperature 1.78482E-05 0.000245526 0.046048966 Humidity 3.39572E-04 0.033262784 0.095357978 Sensor height 0.1086828568 0.639292914 0.3578353637 Low p-values except for CO, 03, and height
- 12. Change in Gases vs # of Humans 12 CO2 vs # of humans present SO2 vs # of humans present Change in CO2 and SO2 due to human entrance or exit is sometimes visually understandable
- 13. Change in Gases vs # of Humans (contd.) 13 LPG vs # of humans present NO2 vs # of humans present Change in LPG and NO2 due to human entrance or exit is NOT generally visually understandable
- 14. Change in Environmental Parameters vs # of Humans 14 Humidity vs # of humans present Temperature vs # of humans present Change in humidity and temperature due to human entrance or exit is NOT generally visually understandable
- 15. Application of Machine Learning Algorithms 15 Classifier Accuracy Room-1 Room-2 Room-3 SMO 91% 49% 48% HMM 9% 9% 8% Classifier Accuracy Room-1 Room-2 Room-3 IBK 99.9% 70.1% 69.2% Random Forest 99.9% 84.2% 85.9% J48 99.5% 83.7% 85.5% Bagging 99.5% 85.4% 85.4% REPTree 99.3% 86.7% 84.5% NaiveBayes 88.4% 53.1% 51.1% DecisionStump 88.1% 46.8% 46.6% Sequential Classifiers Non-Sequential Classifiers Non-sequential classifiers show higher accuracy
- 16. A Potential Reason behind Bad Performance of Sequential Classifier: Noise in Data 16 Room-1 Room-2 Room-3
- 17. Application of Machine Learning Algorithms on Data Sets Filtered using PCA 17 Classifier Accuracy Room-1 Room-2 Room-3 SMO 90% 49% 48% HMM 9% 10% 15% Classifier Accuracy Room-1 Room-2 Room-3 IBK 99.9% 70.1% 69.1% Random Forest 99.9% 83.8% 81.3% J48 99.5% 83.7% 83.2% Bagging 99.6% 85.4% 83.2% REPTree 99.2% 86.7% 80.9% NaiveBayes 89.6% 53.4% 52.1% DecisionStump 88.1% 47.1% 46.2% Sequential Classifiers Non Sequential Classifiers Accuracy for filtered data is roughly same to unfiltered data
- 18. Real-Time Human Counter 18 Classifier Precision Recall IBK 0.990 0.990 Random Forest 0.991 0.991 J48 0.990 0.990 Bagging 0.990 0.990 REPTree 0.991 0.991 NaiveBayes 0.880 0.901 DecisionStump 0.783 0.885 Tradeoff between accuracy and energy through tuning uploading interval 37.5 seconds gives the best trade-off Accuracy is >99%
- 19. Conclusion and Future Work • Human counting based on environmental sensing is little explored in the literature • We propose an ML-based human counting technique through environmental sensing in closed indoor settings – Accuracy is >99% – Can be hidden in a secured place, which can offer a great remedy to prevent any intrusion in a secured indoor space • Future work – Find other possible environmental factors that can influence the human presence – Explore counting human presence in outdoor environment based on environmental parameters 19
- 20. Thank you Questions are welcome! Email: [email protected] 20
- 21. Existing Research Studies • Human Tracking System Based on PIR Sensor Network and Video – Xiong et al., CWSN, 2014 – Presents a tracking algorithm based on pyroelectric sensor network and video analysis technologies to detect and locate the human target motion precisely • Human Movement Detection and Identification Using Pyroelectric Infrared Sensors – Yun et al., Sensors, 2014 – Uses a set of PIR sensors to present an empirical study of human movement detection and identification 21 Existing sensors-based systems suffers from a risk of being deactivated by an intruder due to its visibility
- 22. Existing Human Counting Techniques (contd.) • Human Sensing Using Visible Light Communication – Li et al., MobiCom, 2015 – Presents a new system named as LiSense that can produce real-time human skeleton reconstruction using Visible Light Communication (VLC) • An information technology enabled sustainability test-bed (ITEST) for occupancy detection through an environmental sensing network – Dong et al., Energy and Buildings, 2010 – Describes a large-scale wireless and wired environmental sensor network test-bed and its application to occupancy detection in an open-plan office building 22 Existing sensors-based systems cannot be used in highly secured places as they can be damaged by the intruder
Editor's Notes
- #6: Presents a tracking algorithm based on pyroelectric sensor network and video analysis technologies to detect and locate the human target motion precisely CO2, carbon-monoxide (CO), total volatile organic compounds (TVOC), small particulates (PM2.5), acoustics, illumination, motion, temperature, and humidity. – Deals with uncertainty in open spaces
- #7: We propose a secured human counting technique through sensing four environmental gaseous parameters (CO2, LPG, NO2, and SO2) and two weather parameters (temperature and humidity) in a closed indoor settings
- #10: Data collection period????
- #16: Sequential minimal optimization (SMO)
- #19: Data collection interval???