Improving the safety of ride hailing services using iot analytics
- 1. Improving the Safety of Ride Hailing Services - Using IOT Analytics
- 2. Project Overview The Industry • Ride-hailing Taxi Service Market to Garner $126.52 Billion by 2025 at 16.5% CAGR. • Primary reason for such a growth rate being rising trend of on- demand transportation services, high-end employment opportunities and lower rate of car ownership among millennials The Challenge • Even since the emergence of these Ride hailing companies, there has been an increase in 3 to 5 % of accidents. • Many Research concludes that fleet or company drivers have an increased crash risk relative to that of privately registered vehicle drivers. The Objective • Analyze and capture driving behaviors that are hazardous • Develop a predictive model for predicting Unsafe Trips • Improve the over all Safety of the ride hailing services
- 3. IOT & Analytics – a Powerful Combination • Internet of Things (IOT) – • By 2025, there will be 116 million IOT enabled cars in the U.S. And each connected car will upload 25 GB of data per hour (~ 219 TB / Yr.) • IOT – a costly piece of technology, is it? • Well, no it isn’t anymore • All average smartphones today are equipped with basic embedded low-cost telemetry sensors such as accelerometers, gyroscopes, GPS etc. • Role of IOT Analytics • Large sets of driving data from GPS and telemetry sensors allows for exciting new research possibilities using advance Analytics and Machine Learning techniques • Insights from IOT Analytics help address key concerns facing industries
- 4. IOT Sensors – Accelerometer & Gyroscope • What is an Accelerometer? • Accelerometer sensor reports the acceleration of the device along the 3 sensor axes (X, Y, Z) • The measured acceleration includes both the physical acceleration (change of velocity) and the gravity • All values are in SI units (m/s^2) • What is a Gyroscope? • A gyroscope sensor reports the rate of rotation of the device around the 3 sensor axes (X, Y, Z) • Rotations can be of 3 different types • Pitch – Pitch is for Y axis rotational rate in (rad/s) • Roll – Roll is for X axis rotational rate in (rad/s) • Yaw – Yaw is for Z axis rotational rate in (rad/s)
- 6. Exploratory Data Analysis Acceleration X, Acceleration Z, Gyro Y show a significant difference when the trip is classified as Unsafe. These are initial pointers that these features may play significant role in determining the trip type
- 7. EDA (Contd.) • Speed • Lower speed bins have higher percentage of unsafe rides. Reasons could be • Driving below speed limits • Talking or Texting while driving • Duration • Longer durations Trips have higher percentage of unsafe rides. Reason could be • Drowsy driving due to longer driving hours • Customers tendency to rate longer drivers as uncomfortable • Clearly both Speed and Trip Duration have a significant role in classifying a Trip Quality
- 8. Telemetry Data Processing Methods • Removing Noise component from Raw Accelerometer and Gyroscope Data using Low Pass Filter alpha = 0.8 Filter BEGIN Low_X = alpha × Prev_Low_X + (1 − alpha) × Curr_X Low_Y = alpha × Prev_Low_Y + (1 − alpha) × Curr_Y Low_Z = alpha × Prev_Low_Z + (1 − alpha) × Curr_Z Filter END • Removing Gravity component from the Accelerometer data 𝑎ℎ𝑜𝑟 = 𝐴𝑐𝑐𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 − 𝐴𝑐𝑐𝑔 Where 𝐴𝑐𝑐𝑔 is the average of acceleration over a window size of 100 seconds • Calculate Magnitude or Total Acceleration / Total Angular Velocity 𝑎𝑐𝑐𝑅 = 𝑎𝑥2 + 𝑎𝑦2 + 𝑎𝑧2 𝑎𝑐𝑐𝐻 = 𝑎𝑥ℎ𝑜𝑟 2 + 𝑎𝑦ℎ𝑜𝑟 2 + 𝑎𝑧ℎ𝑜𝑟 2 𝑣𝑒𝑙𝑅 = 𝑣𝑥2 + 𝑣𝑦2 + 𝑣𝑧2 𝑣𝑒𝑙𝐹 = 𝑣𝑥𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 2 + 𝑣𝑦𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 2 + 𝑣𝑧𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 2 Where, ax, ay, az are raw accelerations along x, y, z axis respectively 𝑎𝑥ℎ𝑜𝑟 , 𝑎𝑦ℎ𝑜𝑟 , 𝑎𝑧ℎ𝑜𝑟 are filtered accelerations with gravity treatment done vx, vy, vz are raw angular velocity along x, y, z axis respectively 𝑣𝑥𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 , 𝑣𝑦𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 . 𝑣𝑧𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑 are filtered angular velocity Data Processing (2) Noise Treatment • (2A) Treating Raw data through LPF Gravity Treatment • (2B) Remove Gravity from Accelerometer Raw data Calculate Magnitude • (2C) Calculate Magnitude of Acceleration and of Angular Velocity
- 9. Raw Vs. Filtered Telemetry Data
- 10. Significant Driving Events and Patterns • STEP 1 – Aggressive Turn Events • Turn events is based on filtered gyroscope energy (𝑣𝑒𝑙𝐹) • velF >= 0.025 • STEP 2 – Aggressive Acceleration / Breaking Events • Acceleration (braking/accelerating) events identification, from vehicle acceleration component information (accH) • accH >= 0.002 • STEP 3 – Zig-Zag Events • Zig-zag events are identified as two or more change lanes from the significant measurements of the accelerometer (accH) with very less angular Velocity (velR) • velF <= 0.0025 & accH >= 0.0013 • STEP 4 – Normal Events • Any instance that are not classified as one of the above 3 are classified as normal events Thresholds For Event Identification Thresholds are identified using a combination of mean, Outliers and trial and error method
- 11. Predictive Modeling & Validation Model Precision Recall F1-Score Accuracy AUC GNB 0.46 0.28 0.35 0.7483 0.656 CART 0.34 0.36 0.35 0.6781 0.569 LOGIT 0.70 0.15 0.24 0.7810 0.683 RF 0.48 0.20 0.29 0.7576 0.627 KNN 0.42 0.25 0.31 0.7387 0.618 ANN 0.75 0.14 0.23 0.7828 0.691 XGB 0.75 0.15 0.25 0.7844 0.691
- 12. Conclusion & Recommendations Predicted Unsafe Rides • Alerted on real time basis to respective drivers giving them opportunity to correct it • Alert the drivers on driving behaviors that are Unsafe Performance Evaluation • Drivers can be evaluated based on their driving behaviors thus encouraging better driving practices among drivers Organizing Mandatory Trainings • With specific importance on Feature’s relative importance from Models • Going slower that designated speed limits seems to have caused more Trips unsafe • Higher Trip Duration tend to be more unsafe. Encourage drivers to take short / traffic less routes • Advised to avoid Zig-Zag movements as they cause more trips unsafe
- 13. Appendix