Crash detector
summary
A crash detector uses a CNN on multivariate time series (time-series treated as images) to filter ~200,000 daily signals down to the ~100 real accidents per day problem, and runs in production with automated retraining and validation. Training pipeline uses BigQuery as data lake, Dataproc PySpark for feature extraction and Vertex AI pipelines to train and register TF models, exporting TF Lite for inference; inference runs on Pub/Sub + Cloud Function. Explainability uses SHAP, shapelets and custom multivariate shapelets; system reduces false positives and is deployed via Terraform.
AI-written summary · about
chapters (8)
- 0:00Introduction to crash detector
- 1:07Black box and emergency calls
- 6:01Imbalanced data challenge
- 8:27Neural network approach
- 10:11Explainability in AI models
- 14:53Production deployment overview
- 21:22Model training architecture
- 23:11Q&A session
related talks
40mBuilding Real-Time Applications: Cyclist Crash Detection
23mMaking Sense of Streaming Sensor Data: How Uber Detects On-trip Car Crashes
38mHow We Built a Safety Violations Detection System
38mSampling and Modelling the World in Real-Time for Mobility
22mBuilding a movement data and analytics platform to de-stress customers & save lives
25mMonitoring AI with AI
24mHunting Criminals with Hybrid Analytics
25mReal-Time Fraud Detection at Scale