Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Nericell: Rich Monitoring of Road and Traffic Conditions using Mobile SmartphonesPrashanth Mohan, Venkata Padmanabhan, Ramachandran Ramjee. Microsoft Research India, BangalorePresented by: Philip ShiblyTopics●What is Nericell?●Framework●Acceleration●Orientation and Reorientation●Validation●Road and Traffic Conditions–Bump Detection–Stop-and-Go–Audio●Localization●EnergyWhat is Nericell?●Smartphone system that monitors road and conditions in developing regions.●Tailored to complex traffic flow and degraded roadways.●Avoids usage of infrastructure or area networks.● Tested in Bangalore, India.Framework●How to sense congestion?●How to sense road conditions?●How to determine location?●How to be efficient?Congestion: accelerometer●Monitors acceleration to determine braking and congestion.●Uses the 3-axis accelerometer but must account for orientation.–Does not make assumption that everyone has their phone in the same orientation in the car.–Developed an algorithm for determining orientation and then virtually reorienting the accelerometer.–Monitors for user-interaction with the phone and neglects those readings.●Validates the reorientation and develops heuristics to detect bumps, potholes and braking under certain speed conditions.●Defines orthogonal axes with respect to the phone as (x, y, z).●Defines orthogonal axes with respect to the vehicle as (X, Y, Z).●If they are equal, we say that the phone is well-oriented. If not, it is disoriented.●Accelerometer readings relate to the frame of references as––If the accelerometer is well-oriented, these are equal.●A DC accelerometer is capable of measuring static acceleration as 1G in the downward direction.Acceleration(ax, ay, az)and (aX, aY, aZ)Determining Orientation●Applying rotations about the X, Y, and Z axes are computationally intensive and require the CPU to perform costly trig functions.●Instead, Euler angles are used and any orientation of the accelerometer can be represented by a pre-rotation of about the Z axis, a tilt of about the Y axis, and then a post-rotation of again about the Z axis. ΦpreΘtiltΨpostEstimating Pre-Rotation and Tilt●When the accelerometer is stationary, the only effect on it is gravity (1G) along Z. By their math, this is the only sampling that is needed to calculate pre-rotation and tilt.●Instead of waiting for the vehicle to come to a stop in order to estimate, a rolling 10-second averaging window of the accelerations are taken to determine the median values. Since any momentary bumps would average out, then the pre-rotation and tilt would be able to be continuously calculated.Estimating Post-Rotation●Braking provides a significant change in acceleration that is orthogonal to Z, which is needed to estimate post-rotation.●However, GPS is needed to sample this change and is expensive compared to the prior estimations.●So they monitor pre-rotation and tilt for any noticeable change, then turn GPS on to estimate post-rotation.Validating the Estimations●How successful is reorientation?●A well-oriented accelerometer is compared to a disoriented one and the cross-correlation is taken to determine the effectiveness of reorientation. The cross-correlation is then also compared against the cross-correlation of two well-adjusted accelerometers.Inferring Road and Traffic Conditions●Brake Detection is performed by monitoring●The mean is computed over a sliding window N seconds wide and if the mean exceeds a threshold T, then a braking event is detected.● To establish the ground truth, GPS or CAN is used and a threshold of 1m/s^2 is over a duration of 4 seconds is used to quantify a braking event.●By using T=.11G-.12G in experiments, that equates to 10-20% more conservative than the ground truth.axBrake Detection Results●ACL1 and ACL3 agree quite well●False-positives seem high, but they correlate to deceleration events at a slightly lower rate than the heuristic. ●False-negatives are low, and fall within the GPS localization error. These can be avoided by using the CAN.False Negatives False PositivesAccelerometer (threshold T (g))Rate Change in Speed avg(max)Rate Change in Speed avg(max)ACL-1 (T=0.11) ACL-1 (T=0.12)4.4% 11.1% 15(16)16(18)22.2%15.5%12(10)12(9)ACL-3 (T=0.11) ACL-3 (T=0.12)4.4% 11.1% 15(16)16(18)31.1%17.7%12(9)12(9)Differentiating Between Stop-and-Go and Pedestrian Traffic●Main observations are:– the amplitude of the surges in acceleration–The frequency of the surges●No false positives or negatives were captured using the same heuristics from prior experiment.●Different pedestrian traces did not produce any false positives.Bump Detection●Difficult to implement for a number of reasons–How do you establish the ground truth?●Manual annotation–Accelerometer signal is of very short duration and different magnitude at different speeds.●Implement two sets of heuristics for fast and slow speed.–z_sustaining looks for longer duration dips below a lower threshold.–z_peak looks for quicker dips below a higher threshold–Needs a training set to develop heuristicszsusand zpeakBump Detection Results Low Speed High SpeedBump Detection Results●Both detectors tuned for low false-positives (<10%)●False-negatives are high because of the difficulty of establishing the ground truth.Honk Detection●Implements a simple heuristic approach for both exposed and enclosed vehicles. ●Detector implements a discrete Fourier transform on 100ms audio samples and looks at spikes in the frequency domain. ●By observation, they choose to implement the heuristic by looking for two spikes in frequency with one required to be within 2.5kHz to 4.0kHz range. ●Ground truth is by manual annotation.Honk Detection Results●With a large enough spike to avoid false-positives, the detector performs better in exposed vehicles due to a higher received power.●The varying sensitivity between phones produces a different number of honks detected within the same sample. ●A high spike threshold protects against false-positives but doesnt eliminate them.Localization●GPS and WiFi are high energy users. ●GSM is much more efficient when using signal strength localization algorithms. Although, it relies on dense network of towers that is
View Full Document
Unlocking...