Advertisement
Advertisement

More Related Content

Similar to [Ubicomp'15]SakuraSensor: Quasi-Realtime Cherry-Lined Roads Detection through Participatory Video Sensing by Cars(20)

Advertisement

More from Ubi NAIST(20)

Advertisement

[Ubicomp'15]SakuraSensor: Quasi-Realtime Cherry-Lined Roads Detection through Participatory Video Sensing by Cars

  1. SakuraSensor: Quasi-Realtime Cherry-Lined Roads Detection through Participatory Video Sensing by Cars Shigeya Morishita†, Shogo Maenaka†, Daichi Nagata† Morihiko Tamai†, Keiichi Yasumoto†, Toshinobu Fukukura‡, Keita Sato‡ †Nara Institute of Science and Technology ‡DENSO CORPORATION
  2. Latest car navigation systems • Help drivers search comfortable & efficient routes • Criteria – Traveling distance – Traveling time – Toll/Toll-free – Fuel efficiency – Scenic beauty 2 © NAVITIME (http://products.navitime.co.jp/function/2519.html」) Toll-free Fuel-efficient Minimum distance Toll Scenic
  3. Scenic route search Problems of existing services • Information is edited manually – Small number of scenic spots – Low update frequency • Scenery information consists of only texts and images – insufficient for users 3 Our approach • Use participatory sensing by cars • Collect and share videos of scenic spots Example of scenic spot info.
  4. Related work 4 Method Proposed method ParkNet [12] SignalGuru [15] Nericell [3] Participatory sensing ○ ○ Cooperative sensing ○ Real-time ○ ○ ○ ○ Information detection from videos ○ × (ultrasound signals) △(traffic signals) × (horn sounds) [12] ParkNet: Drive-by Sensing of Road-Side Parking Statistics, MobiSys’10 [15] SignalGuru: Leveraging Mobile Phones for Collaborative Traffic Signal Schedule Advisory, MobiSys’11 [11] Nericell: Rich Monitoring of Road and Traffic Conditions using Mobile Smartphones, SenSys’08 Many existing studies on participatory sensing (PS) by cars No studies use both PS and real-time video sensing
  5. SakuraSensor: automatically identifies scenic spots location and collects videos using PS ・ we target cherry-lined roads ・ automatically collect and update scenic information ・ gathering videos of scenic location The best period of flowering cherries is short and uncertain from year to year and from place to place
  6. SakuraSensor App for iOS devices 6Full size video - https://youtu.be/2pRfDS7DeAc Demo at Hall C No.20
  7. Key Idea 7 CloudCars with Smartphone Too much cost for cellular bandwidth & computation resource at cloud Recording video Analyzing & sharing video with cherriesUpload whole recorded video Recording video Analyzing video Upload only video with flowering cherries Sharing video with cherries
  8. Technical Challenges TC1: Real-Time flowering cherry detection by smartphone TC2: Efficient load distribution among cars 8
  9. TC1: Real-time Cherry Detection • Employ simple computer vision techniques – Smart phone has lower computation power than PC/Cloud Basic approach • Count cherry-like color pixels in each image • Identify amount of flowering cherry as cherry intensity Problem to solve • Artificial objects with similar color must be removed 9
  10. Step1: Removing Artificial Objects An input image Binary image after edge detection box counting method [5] fractal dimensions10 • Employ fractal analysis – Note: natural objects has higher fractal dimension
  11. Real-time fractal dimension calculation 11 Red regions show natural objects
  12. Step2: Detecting Cherry by Color Analysis 12 Used 148 regions extracted from various scenes • Created color histogram of flowering cherry in HSV color space
  13. HSV color space • H(Hue) • S(Saturation) • V(Value of Brightness) From 「http://en.wikipedia.org/wiki/HSL_and_HSV」 characterizes the color significantly varies depending on the lighting condition 13 Our approach used only H-S color space
  14. H-S histogram for flowering cherry H S 0 179 0 255 • Created from total of 148 cherry regions • The value at each coordinate is normalized between 0 and 1 14
  15. Step3: Calculating cherry intensity of an image H S 0 0 Pixel’s (H, S)=(30, 20) The value of (30, 20) is 0.816 An input image Cherry intensity = average value of all pixels Use Backprojection method [6]
  16. Real-time cherry intensity calculation 16 Red boxes show high cherry intensity regions
  17. TC2: Load Distribution among Cars When all cars always conduct image analysis & uploads too much cost (battery consumption, bandwidth, etc) Possible approach • each car senses at a fixed interval may miss PoI (cherry locations) 17
  18. k-stage sensing 18 location where sensing is performed Narrows sensing interval step-by-step when new PoI is found Fixed interval (1st stage) PoI is detected! The preceding car
  19. k-stage sensing 19 Shorter Interval (2nd stage) PoI is detected! Sensing is performed in this Radius PoI detected by preceding car The following car traveling the same road Narrows sensing interval step-by-step when new PoI is found location where sensing is performed
  20. Evaluation of SakuraSensor • Investigate effectiveness of cherry intensity – Compare the results of manual classification and automatic classification by cherry intensity Videos manual classification (used as ground truth) classification by cherry intensity Compute accuracy by comparison 20
  21. Videos used for experiments • Recorded videos in 8 different scenes (routes) using SakuraSensor app for iOS by multiple cars scene name date vehicle area Length (min.) S1 Mar. 31 V1 Aichi Pref. 17 S2 Apr. 5 V2 Nara Pref. 12 S3 Apr. 10 V2 Nara Pref. 66 S4 Apr. 10 V3 Nara Pref. 261 S5 Apr. 10 V4 Nara Pref. 186 S6 Apr. 11 V1 Gifu Pref. 72 S7 Apr. 12 V2 Osaka Pref. 137 S8 Apr. 18 V1 Aichi Pref. 89 extracted 1-second videos at random starting time from each scene 21
  22. 1-Second Videos Manual Classification Class name Criteria C1 cherry ratio (in image) < 5% C2 5% ≤ cherry ratio < 25% C3 25% ≤ cherry ratio Scene C1 C2 C3 S1 79 17 10 S2 93 10 17 S3 372 43 3 S4 1613 96 45 S5 1167 6 0 S6 261 47 72 S7 888 1 0 S8 521 10 7 Total 4994 230 154 22 • Classification results with the same decision by two persons were used
  23. Videos of each class 23 C1 (ratio < 5%) C2 (5% ≤ ratio < 25%) C3 (25% ≤ ratio)
  24. Evaluation Methodology 𝐶1 𝐶1 𝐶2 𝐶2 𝐶3 𝐶3 Dividing videos of each class into halves Training set Test set 24 Set of 1 second videos Set of videos of class 𝐶1 Set of videos of 𝐶2 Set of videos of 𝐶3 Manual classification by human
  25. Evaluation Methodology 𝐶1 𝐶1 𝐶2 𝐶2 𝐶3 𝐶3 Training set Test set Median of cherry intensity: M1 (0.00033) Median of cherry intensity: M2 (0.00791) Median of cherry intensity: M3 (0.03326) Vi Cherry intensity A video 25 𝐷(𝑉𝑖) 𝑀1 𝑀2 𝑀3 𝑉𝑖 is classified into the class that has smallest distance between 𝐷 𝑉𝑖 and its median
  26. Classification Accuracy (1-second videos) • 𝑪 𝟏 and 𝑪 𝟑: good results • 𝑪 𝟐: not enough 26 𝐶1 𝐶2 𝐶3 precision recall 0.97 0.90 0.74 0.83 0.24 0.65
  27. Evaluation of k-stage sensing 27 • Simulation by 600 cars (k=3, 300m150m50m) smaller sensing times similar PoI discovery rate
  28. Conclusions • SakuraSensor – Participatory video sensing system by cars – Consisting of two key techniques • Flowering cherry detection by in-vehicle smartphone – Color histogram analysis for identifying cherry-blossoms – Fractal dimension analysis for removing artificial objects other than flowering cherry – Cherry detection accuracy (C3) with 0.7 of Precision and 0.8 of Recall • k-stage sensing – Distribute sensing load among cars – Similar PoI discovery rate with about half sensing times compared with the fixed interval sensing method 28
  29. 29 Thank you! Demonstration at Hall C No.20

Editor's Notes

  1. Thank you chairperson. Good afternoon, everyone. My name is Shigeya Morishita from Nara Institute of Science and Technology. I am very happy to see all of you today. Today, I would like to present our research named sakura sensor.
  2. Latest car navigation systems help drivers with comfortable and efficient driving. With these systems, we can search routes by various (ベアリアス) criteria. Among these criteria, we focus on scenic beauty.
  3. However, existing scenic route search services have some problems. First, information is edited manually. Second, scenery information consists of only texts and images. To solve these problems, approach, we use participatory sensing by cars and automatically collect and share videos (大きな声で強調) of scenic spots.
  4. There are many existing studies on participatory sensing by cars. However, as long as we know, no studies use both participatory sensing and real-time video sensing. 提案手法の新規性や工夫について.
  5. We propose Sakura Sensor, which automatically identifies scenic spots location and collects videos using participatory sensing. SakuraSensor targets flowering cherries called SAKURA in Japanese, since the best period of flowering cherries is short and uncertain from year to year and from place to place. So, up-to-date information is mandatory.
  6. We have developed SakuraSensor application for iOS devices. I’ll show a demo video of sakurasensor. We are also demonstrating SakuraSensor at hall C number twenty.
  7. One possible approach to realize SakuraSensor is as follows. Cars with smartphone record videos and upload the whole recorded video to cloud server for analysis and sharing. However, this approach takes too much cost for cellular bandwidth and computation resource at cloud. The key idea of SakuraSensor is analyzing video at smartphones so that only video with flowering cherries are uploaded to the cloud and shared. Challenges and key ideas of SakuraSensor
  8. We have two technical challenges to realize SakuraSensor. First challenge is how to realize real-time flowering cherry detection by smartphone. Second challenge is how to realize efficient load distribution among cars.
  9. For the first challenge, we employ simple computer vision techniques since smart phone has lower computation power. So, our basic approach is just to count chery-like color pixels in each image and identify amount of flowering cherry In each image called cherry intensity (強調). Here, the problem to solve is that artificial objects with similar color must be removed.
  10. To remove artificial objects in each image, we employ fractal dimension analysis. Here, note that natural objects has higher fractal dimension. So, to an input image, we apply edge detection algorithm, and box counting method to calculate fractal dimension of each square region.
  11. This is Real-time fractal dimension calculation. Here red color regions show natural objects.
  12. Then we detect flowering cherry by color analysis. We created color histogram of flowering cherry in HSV color space. Here, we used 148 regions extracted from various scenes. These are part of the regions.
  13. HSV color space consists of Hue, Saturation and Value of brightness. From preliminary experiment, we found that V significantly varies (ヴェアリーズ) depending on the lighting condition. So, we used only H-S color space.
  14. This is the H-S histogram created from 148 cherry regions. The value at each coordinate is normalized between 0 and 1.
  15. Then we calculate cherry intensity of an image by using backprojection method. For each pixel, the value is retrieved in the H-S histogram Finally, cherry intensity of the image is calculated as the average value of all pixels.
  16. This is real-time cherry intensity calculation. Here, red boxes show high cherry intensity regions.
  17. The second technical challenge is load distribution among cars. When all cars always conduct image analysis and uploads of videos, the cars will take too much cost. Possible approach is that each car senses at a fixed interval. However, it may miss PoI.
  18. So we propose k-stage sensing which narrows sensing interval step-by-step when new PoI is found by preceding cars. This is the example of k-stage sensing. The preceding car travels and sensing is performed at an initial fixed interval.
  19. After that, when a following car enters the same road. The car narrows its sensing interval and radius, respectively, because a PoI is found on the road. Then, this car performs sensing at the shorter interval while the car is in the circle centered at the PoI with radius.
  20. We conducted some experiments to evaluate Sakura Sensor. The first experiment is to investigate the accuracy of cherry intensity. We compare the result of manual classification and automatic classification by cherry intensity.
  21. We recorded videos in 8 different scenes using SakuraSensor application for iOS devices by multiple cars. We extracted 1-second videos at randomly selected starting time from each scene.
  22. We defined three classes where C1’s cherry ratio in image is less than 5%, C2 between 5 and 25%, C3 more than 25%. Here, only classification results with the same decision by two persons were used.
  23. These are example Videos of class C1
  24. First, we divided the set of classified videos in each class to the training set and the test set.
  25. Then, from training set, we calculated median of cherry intensity for each class. Using the median values, 1-second videos in the test set are automatically classified.
  26. This figure shows the classification result by Sakura Sensor. We see that a good classification result is obtained for class C1 and C3 videos. On the other hand, for class C2 videos result is not so good. The main reason is that many videos included in class C1 were classified to class C2.
  27. We also evaluated the effectiveness of 3-stage sensing method. Evaluation was done with simulation by 600 cars. These results show the k-stage sensing achieves good PoI discovery rate with smaller sensing times. We just adopted hulistic. Or empilically
  28. Conclusions We proposed sakura sensor which is a Participatory video sensing system by cars. As two key techniques, we proposed flowering cherry detection by in-vehicle smart phone and K-stage sensing. (以降は話さない) this system consisting of two key techniques First is flowering cherry detection by in-vehicle smartphone. Color histogram analysis for identifying cherry-blossoms Fractal dimension analysis for removing artificial objects other than flowering cherry Cherry detection accuracy with 0.7 of Precision and 0.8 of Recall Second is k-stage sensing this method distributes sensing load among cars. Similar PoI discovery rate with about half sensing times compared with the fixed interval sensing method.
  29. We also demonstrate our system. Please wat どれだけリアルタイムで桜センサの計算ができるかという説明のスライドを追加 実装上の工夫
Advertisement