Animal Monitoring with Unmanned Aerial Vehicle-Aided Wireless Sensor Networks Jun Xu, G¨ urkan Solmaz, Rouhollah Rahmatizadeh, Damla Turgut and Ladislau B¨ ol¨ oni Department of Electrical Engineering and Computer Science University of Central Florida - Orlando, FL October 24, 2015 Jun Xu (UCF) LCN 2015 October 24, 2015 1 / 21
Outline Introduction 1 Motivation & problem statement Jun Xu (UCF) LCN 2015 October 24, 2015 2 / 21
Outline Introduction 1 Motivation & problem statement Problem analysis 2 Clustering Network modeling Value of information Jun Xu (UCF) LCN 2015 October 24, 2015 2 / 21
Outline Introduction 1 Motivation & problem statement Problem analysis 2 Clustering Network modeling Value of information Proposed path planning approach 3 Markov decision process Path planning process for UAV Jun Xu (UCF) LCN 2015 October 24, 2015 2 / 21
Outline Introduction 1 Motivation & problem statement Problem analysis 2 Clustering Network modeling Value of information Proposed path planning approach 3 Markov decision process Path planning process for UAV Simulation study 4 Simulation setup and metrics Demo display Performance results Jun Xu (UCF) LCN 2015 October 24, 2015 2 / 21
Outline Introduction 1 Motivation & problem statement Problem analysis 2 Clustering Network modeling Value of information Proposed path planning approach 3 Markov decision process Path planning process for UAV Simulation study 4 Simulation setup and metrics Demo display Performance results Conclusion 5 Jun Xu (UCF) LCN 2015 October 24, 2015 2 / 21
Motivation Animal monitoring have various goals: Tracking their migration paths Predict if specific endangered species exist Goal of this application: Providing reliable animal appearance information in large-scale areas Do not using mounting devices & not affecting animal activities Jun Xu (UCF) LCN 2015 October 24, 2015 3 / 21
Problem statement How to find these animals? Sensors can not directly send data to remote base station How the sink (UAV) knows which sensors have the relevant information How to use those sensed information? Latency between animal appearance and information being gathered How to quantify this information Jun Xu (UCF) LCN 2015 October 24, 2015 4 / 21
Clustering Real movement trajectories of 4 zebras in 3 days † Wildlife animals are more likely to having activities in a small area † Yong Wang, Pei Zhang, Ting Liu, Chris Sadler, Margaret Martonosi, CRAWDAD dataset princetonzebranet (v. 02/14/2007), traceset: movement, downloaded from http://crawdad.org/princeton/zebranet/20070214/movement, doi:10.15783/C77C78, Feb 2007. Jun Xu (UCF) LCN 2015 October 24, 2015 5 / 21
Clustering Real movement trajectories of 4 zebras in 3 days † Wildlife animals are more likely to having activities in a small area † Yong Wang, Pei Zhang, Ting Liu, Chris Sadler, Margaret Martonosi, CRAWDAD dataset princetonzebranet (v. 02/14/2007), traceset: movement, downloaded from http://crawdad.org/princeton/zebranet/20070214/movement, doi:10.15783/C77C78, Feb 2007. Jun Xu (UCF) LCN 2015 October 24, 2015 6 / 21
Network model Cluster-heads are responsible for receiving data from other sensors and submitting data to the UAV Single UAV communicates with cluster-heads Jun Xu (UCF) LCN 2015 October 24, 2015 7 / 21
Value of information The value of information (VoI) † Sensed information has the high- est value when event occurs Our goal is maximizing the VoI in the whole network A: the initial value of the information B: the decay speed of the VoI † Turgut, Damla, and Ladislau B¨ ol¨ oni. ”A pragmatic value-of-information approach for intruder tracking sensor networks.” Communications (ICC), 2012 IEEE International Conference on. IEEE, 2012. Jun Xu (UCF) LCN 2015 October 24, 2015 8 / 21
Markov decision process model 5-tuple { S , A , P , R , γ } : S is the set of states (grids) in the network A is the set of possible actions that UAV can do P is the state transition probabil- ities R is the instant reward when the UAV enters one gird γ ∈ [0 , 1) is the discount param- eter Jun Xu (UCF) LCN 2015 October 24, 2015 9 / 21
Markov decision process model Solved this MDP model by Q-learning 9 possible actions of S 4 : 8 neighbors and staying itself Q ( s , a ) = R ( s ) + a ′ Q ( s ′ , a ′ ) γ max Instant reward R ( s ), future potential re- ward Q ( s ′ , a ′ ) Possible actions of S 4 : { Northwest, North, Northeast, West, Stay, East, Southwest, South, Southeast } Jun Xu (UCF) LCN 2015 October 24, 2015 10 / 21
Path planning flow chart Exploitation: deter- ministic grid selection by Q ( s , a ) value Exploration: random grid selection ǫ : random selection probability Jun Xu (UCF) LCN 2015 October 24, 2015 11 / 21
Simulation setup Movement traces of zebras: ◮ ZebraNet project † ◮ 5 zebras in June 2005 at a 10 km × 10 km area near Nanyuki, Kenya ◮ 5682 GPS records in total ◮ GPS sampling time interval: 1 minute Definition of sensing events: ◮ If zebra switches grid, record the event ◮ If zebra always stays in one grid, record every ∆ t time † Yong Wang, Pei Zhang, Ting Liu, Chris Sadler, Margaret Martonosi, CRAWDAD dataset princetonzebranet (v. 02/14/2007), traceset: movement, downloaded from http://crawdad.org/princeton/zebranet/20070214/movement, doi:10.15783/C77C78, Feb 2007. Jun Xu (UCF) LCN 2015 October 24, 2015 12 / 21
Simulation setup Simulator: ◮ Java-based discrete time simulator Performance metrics: ◮ Value of information ◮ Average message delay ◮ Number of zebras encountered Approaches for comparison: ◮ Greedy ◮ Traveling salesman problem ◮ Random Jun Xu (UCF) LCN 2015 October 24, 2015 13 / 21
Simulation setup Network size 10 km × 10 km Number of grids (states) 16 Grid size 2500 m × 2500 m Unit experimental time (round) 10s UAV speed 100 m / round Decay speed of VoI (parameter B) 0.05 Radius r for direct observation 200 m Initial reward IR ( σ , C i , I dist , I duration ) 10.0 (10.0, 1.0, 1.0, 1.0) Jun Xu (UCF) LCN 2015 October 24, 2015 14 / 21
Demo display Jun Xu (UCF) LCN 2015 October 24, 2015 15 / 21
Value of information 1400 MDP Greedy 1200 TSP Random 1000 Value of information MDP (Markov decision process) 800 Greedy (Greedy total number of pre- vious events) 600 TSP (Traveling salesman problem) Random (Random selection from all grids) 400 200 0 0 500 1000 1500 2000 2500 3000 Simulation time (min) Jun Xu (UCF) LCN 2015 October 24, 2015 16 / 21
Average message delays 60 50 40 Average delay (min) MDP (Markov decision process) Greedy (Greedy total number of pre- 30 vious events) TSP (Traveling salesman problem) Random (Random selection from all 20 grids) 10 0 MDP Greedy TSP Random TSP: 0 deviation because fixed route Jun Xu (UCF) LCN 2015 October 24, 2015 17 / 21
Number of zebras encountered 250 200 Number of zebras encountered MDP (Markov decision process) 150 Greedy (Greedy total number of pre- vious events) TSP (Traveling salesman problem) 100 Random (Random selection from all grids) 50 0 MDP Greedy TSP Random Direct observation radius ( r ) Jun Xu (UCF) LCN 2015 October 24, 2015 18 / 21
Performance stability 1500 MDP−1 MDP−2 MDP−3 MDP−4 1000 Value of information Results from 4 time experiments Same parameters 500 0 0 500 1000 1500 2000 2500 3000 Simulation time (min) Jun Xu (UCF) LCN 2015 October 24, 2015 19 / 21
Exploration 1500 ε = 0.2 ε = 0.4 ε = 0.6 ε = 0.8 1000 Value of information ǫ − Greedy policy The probability of random grid selec- tion by UAV 500 0 0 500 1000 1500 2000 2500 3000 Simulation time (min) Impact of exploration Jun Xu (UCF) LCN 2015 October 24, 2015 20 / 21
Conclusion We focused on the animal monitoring in large area We proposed a MDP-based approach for UAV path planning The evaluation indicated significant improvement compared to Greedy, TSP and Random Future work: ◮ Other species, other dataset ◮ Multi-UAVs Jun Xu (UCF) LCN 2015 October 24, 2015 21 / 21
Recommend
More recommend
