How can we use Data Science to measure people's perception of safety
Download as PPTX, PDF0 likes116 views
The document discusses the development of a perceived safety index for urban environments, as presented by Claudia Soares and her collaborators. It outlines the methodology, including the use of pairwise comparisons and geo-referenced urban data, to assess how people perceive safety in cities. The project also emphasizes the importance of data integration and innovative techniques to improve the accuracy and reliability of the safety estimates.
![Semantic segmentation with In-PlaceABN+WideResNet38+Deeplabv3 trained on Mappilary Vistas [1]
[1] S. R. Bulò, L. Porzi, and P. Kontschieder. In-Place Activated BatchNorm for Memory-Optimized Training of DNNs. In 2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition, pages 5639-5647. IEEE, 2018
Extracting semantics from GSV images
Sidewalks
People
Crosswalks
cars](https://image.slidesharecdn.com/claudiasoaresdsc-191120145347/85/How-can-we-use-Data-Science-to-measure-people-s-perception-of-safety-14-320.jpg)
How can we use Data Science to measure people's perception of safety
- 1. TITLE SPEAKER’S NAME City-SAFE measuring perceived safety in urban environments Claudia Soares Joint work with Gabriel Costa, Manuel Marques
- 2. Who am I? 2 Creativity Human culture and interaction Communication Mathematical grounds Engineering skills Implementation-oriented Initiative and self-direction Flexibility Productivity Optimization, SW Engineering Team culture Project management and B2B PhD
- 3. Who I am 3 @Data Science made in Switzerland
- 4. 00 10 20 30 40 50 60 70 80 90 100 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020 2025 2030 2035 2040 2045 2050 World Portugal Serbia Why Urban? United Nations, Department of Economic and Social Affairs, Population Division (2014). World Urbanization Prospects: The 2014 Revision, custom data acquired via website. %
- 5. Maslow’s hierarchy of needs (1943) Why perceived safety?
- 6. Estimating Perceived Safety: general idea Estimator Perceived safety Geo-referenced Urban data Estimator Perceived safety Train Geo-referenced Urban data Estimator Perceived safety Train Geo-referenced Urban data Construct Ground Truth
- 7. Data Collection Google Maps API ~5k points
- 8. Pairwise Comparisons Survey https://smartcity.isr.tecnico.ulisboa.pt/CitySAFE/ >19k comparisons
- 9. Why pairwise comparisons? Respondents are more precise Less biases Nevertheless… From Psychology research Data processing is harder: no traditional ML tools Respondents are more accurate Less noise
- 10. How do the survey data look like? (detour)
- 11. Respondents voting is a power law
- 12. Respondents are balanced Most active respondents
- 14. Semantic segmentation with In-PlaceABN+WideResNet38+Deeplabv3 trained on Mappilary Vistas [1] [1] S. R. Bulò, L. Porzi, and P. Kontschieder. In-Place Activated BatchNorm for Memory-Optimized Training of DNNs. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5639-5647. IEEE, 2018 Extracting semantics from GSV images Sidewalks People Crosswalks cars
- 15. Common semantics for urban perceived safety Respondent representation example winning images losing images Agglomerative hierarchical clustering Cosine Distance Complete Linkage
- 16. Do we all perceive the same?
- 17. What makes us feel safe?
- 18. Back to the perceived safety index… (end of detour)
- 19. Estimator Perceived safety Train Geo-referenced Urban data Construct Ground Truth
- 20. Global perceived safety index from pairwise comparisons • Local partial ordering • Issue with comparison noise and small sample: 20k comparisons for 5k images • Need to introduce structure via regularization
- 21. Urban mesh continuity Agreement between comparisons Continuity in space and time Some images might not represent a street view An index is an ordering Some comparisons will contradict each other The perceived safety index we want
- 22. Urban mesh Agreement Creating a ground truth for perceived safety: Expressing what we want with a model
- 23. index vector: Agreement model: penalize disagreement linearly to escape outlier votes Comparison data vector:
- 25. Urban mesh regularization: proximity in walking distance Geographical graph j i
- 26. Up-to-date regularization: Accounting for image age in GSV Time graph j i
- 27. Synthetic Data: robust to comparison noise and number of comparisons # comparisons noise level σ 5000 2 8000 2 20000 2 40000 2 80000 2 # comparisons noise level σ 8000 1 8000 2 8000 3 8000 4
- 28. Error vs. #comparisons: Our CitySAFE index outperforms the standard EmpiricalCDF Estimation error
- 29. Error vs. noise power EmpiricalCDF Estimation error
- 30. Synthetic Ground Truth TrueSkillCitySAFE 8000 comparisons σ = 2 Qualitative Results: it pays off to use additional information
- 31. Great fit on the Lisbon dataset!
- 32. What semantics are in safer images?
- 33. EU + National funding
- 34. • Coordination, data aggregation, code • Mobility, citizen science, data science • Gamification, Human Computer Interaction • Citizen interaction, end user Consortium
- 35. Diagnostics for decision-makers and citizens
- 36. Estimator Perceived safety Train Geo-referenced Urban data Construct Ground Truth Back to our learning problem
- 37. Harness pre-existing data sources To deliver valuable information
- 38. Data Collection Google Maps APIs Twitter API Municipal Lisbon Police Na Minha Rua LX Survey Estimator
- 39. Mining free text Topic discovery Latent Dirichelet allocation (LDA) Categorical data type
- 40. Multiple Kernel Learning with Support Vector Regression
- 41. Main devs for the (near) future • Scalability: MKL + SVR not viable for larger datasets • Text: converting to latent spaces, e.g. BERT • New data sources: YouTube travelling vlogs