A Two-Layer Approach for Energy Efficiency in Mobile Location Sensing Applications Ling-Jyh Chen (Academia Sinica, Taiwan)
Introduction • Mobile location sensing applications (MLSAs) • Exploit Global Positioning System (GPS) technology to facilitate location-based services • MLSA platforms are battery-powered and resource-constrained • Tradeoff between information accuracy and energy efficiency • Two-layer of MLSAs • GPS tracking • Data communication
Previous work (1) • Energy efficiency on the MLSA GPS tracking • Static duty-cycle (SDC) • Turn GPS receivers ON and OFF at regular intervals • Operate in a “ blind ” manner • Dynamic duty-cycle (DDC) • Adjust GPS duty cycles based on events triggered by additional sensors or analytical models (ex: a lookup table of pre-learned radio patterns)
Previous work (2) • Energy efficiency on the data communication • Byte-level compression • Compress MLSA data without considering its intrinsic properties • Compressed data cannot be processed directly without decompression into its raw form • Spatio-temporal compression • Achieve a good compression ratio at the expense of information loss
Our solution: Two-layer solution • Energy efficiency on the MLSA GPS tracking: Event-based GPS Tracking (EBT) • Require minimal prerequisites for extra knowledge • Energy efficiency on the data communication: Inter-Frame Coding (IFC) • Provide lossless compression and allow queries to operate on the compressed data directly • Simple, lightweight, and portable to off-the-shelf smart phones
Two-layer solution • Event-based GPS Tracking (EBT) • Inter-Frame Coding (IFC)
Event-based GPS Tracking (EBT) EBT Hybrid duty-cycle Location estimation scheduling Event-based dynamic duty-cycle (DDC): Turn Detection Uses G-sensor data Approach (TDA) to estimate location Static duty-cycle (SDC)
EBT has two components • Hybrid duty-cycle scheduling • Static duty-cycle (SDC) • Turn on GPS when GPS receiver has been in the OFF mode longer than a pre- defined period (TDC) • Event-based dynamic duty-cycle (DDC) • Turn on GPS when detecting a significant change in the movement pattern • Location estimation • Uses the data of the G-sensor to estimate the location when GPS is off
Hybrid duty-cycle scheduling
Question (1) How to detect a turn? • Turn Detection Approach (TDA) : Significant changes in direction result in significant changes in acceleration data • Sliding standard deviation of acceleration in direction orthogonal to gravity and trajectory’s direction (window size = w) • Target Power Saving Ratio ( α ) • If a sliding standard deviation is on the top (1 - α ) , there is a significant change in acceleration data • Set the sliding standard deviation on the top (1 - α ) as turn detection threshold • Poisson Sampling (rate = λ ) • Random view of acceleration distribution
Question (2) Queue management in TDA • Infinite-Queue Approach (IQA) • Provide a baseline , but infeasible • FIFO Queue Approach(FQA) • Finite queue length (L) • Skewed view of acceleration distribution • Dual-Queue Approach (DQA) One is for standard deviations smaller than or equal to threshold ( FIFO queue length = ⎡ α L ⎤ ) The other one is for standard deviations greater than threshold ( FIFO queue length = L - ⎡ α L ⎤ )
Location estimation: Estimate location when GPS is in the OFF mode • Direction • Obtained from the last successful GPS lock • Displacement • Displacement measurement algorithm (DMA) [6] [6] T. Chen, W. Hu, and R. Sun. Displacement Measurement Algorithm Using Handheld Device with Accelerometer. In Asia-Pacific Conference on Wearable Computing Systems, 2010.
Two-layer solution • Event-based GPS Tracking (EBT) • Inter-Frame Coding (IFC)
Inter-Frame Coding (IFC) • Spatial and temporal offsets are limited to • Object’s mobility • Trajectory’s sampling rate • IFC exploits the spatial and temporal localities of contiguous spatio-temporal data to reduce redundancy
Two types of data points in IFC • I frame: Index data points of a trajectory • O frame: Offsets of the subsequent data points • An I frame is associated with n O frames • n depends on sampling rate , speed , and data compression ratio
Example of IFC (given n = 3) • The second I frame is created because n O frames have been created for the first I frame. • The third I frame is created because the longitude and latitude offsets exceed the maximum offset value.
Upper bound of n • Spatial and temporal offsets are limited to • Maximum value of the latitude and longitude offsets = MAX dist • Maximum value of the time offset = MAX time • Maximum possible speed = V max m/s • Trajectory data collection rate = s data points/s
Compression ratio ψ • The best compression ratio is (Size_O/Size_I) when n approaches infinity, but very large n value is infeasible • Computationally expensive when n is very large: Data query involves two separate database queries • Loss of a single I frame may result in the loss of the original data • Cannot achieve the theoretical compression ratio: Subsequent n points have oversized offset values that cannot be represented by O frames
Evaluation • Evaluation of EBT • Evaluation of IFC
Experimental setup • 50 trips of the TPE-CMS bus system using the VProbe application • Collect smart phone sensory data: GPS trajectories, digital compass directions, and 3-axis accelerations • Platform: Acer Liquid, HTC Magic, Samsung Nexus S, and Sony Ericsson XPERIA X10 phones • Configuration • • Static Duty Cycle (T DC ) = 60 Data sampling rate seconds • GPS: 1 Hz • Queue Length (L) = 1000 samples • Digital compass: 20 Hz • Window Size (w) = 50 samples • 3-axis acceleration: 20 Hz • Results are based on the average performance of 10 simulation
Evaluation (1) Feasibility of using digital compasses • Dataset • Trajectories of 86,607 seconds • 235 turn events are marked manually as ground truth • Results • 795 turn events are reported by the digital compass • 115 turn events are detected correctly: accuracy of turn event detection is 48.94% • 680 events are false-alarms: false positive ratio is 85.53% • Digital compasses are very sensitive to magnetic and electrical fields
Evaluation (2) Evaluation of TDA • EBT’s hit rate is more than 98% when the TDA scheme is used (i.e., α < 1) • Hit rate is approximately 71% when α = 1 (i.e., under the SDC scheme) • There are no significant differences between the hit rates of the IQA, FQA and DQA schemes
Evaluation (3) Power saving ratio achieved with different target power saving ratios with the three queue management schemes • α ’ increases with α , and the IQA and DQA schemes perform better than FQA • The reason is that FQA implements the FIFO queue with a size limit of L = 1, 000, and the selected threshold S thresh is not usually representative of the true distribution of turn events; hence, there is a large number of false alarms
Evaluation (4) Location estimation errors using different target power saving ratios • The distance error increases with α : loss of location accuracy is the the trade-off reduced energy consumption • TDA scheme improves the location accuracy significantly in EBT • The average location estimation error is • About 120m when the TDA scheme is not applied (i.e., α = 0) • About 80m (i.e., a 33% improvement) with the TDA scheme and α = 0.95
Evaluation (5) Power saving ratio achieved with different sampling rates and target power saving ratios EBT achieves a good power saving ratio and detects nearly all the turn events under different values
Evaluation • Evaluation of EBT • Evaluation of IFC
Experimental setup • We implement the IFC scheme using the open-source PostgreSQL database (version 8.4.4) and the PostGIS spatial database extension (version 1.5.1) • Data size • • O frame (10 bytes) I frame (32 bytes) • • Integer data type (4 bytes) for I Point data type (16 bytes) for frame sequence number (i) location information (x and y) • • Short Integer data type (2 bytes) Timestamp data type (8 bytes) for offsets Dx, Dy, and Dz for time information (z) • Integer data type (4 bytes) for sequence numbers (i) and trajectory identifiers (u)
Evaluation (6) Comparison of theoretical compression ratio and compression ratio achieved • Two curves are nearly overlapped completely • Compression ratio is lower than 0.5 after the value of n becomes larger than 4
Evaluation (7) Comparison of the results of the IFC, OPW, TDTR, STTrace, Uniform Sample, and DP schemes • We use an exhaustive set of configurations to observe their Pareto frontiers between the average distance error and the compression ratio • IFC scheme outperforms the other schemes significantly and always achieves the “Pareto optimum” • IFC scheme is lossless • IFC’s distance error is zero despite the different compression ratios achieved
Recommend
More recommend