Open AndreasKostler opened 11 years ago
We can also note what signals/data/algorithms etc that can match or reduce the likelihood of false positives.
ie, Sensor readings staying static after a possible crash increases the probability of a real crash. Can we combine sensors for better measurements - ie, accelerometer and gyro. The sensor analysis issues are related to this.
I guess there is at least two ways of looking at this problem: 1) The Bayesian approach: Gather empirical data (i.e. probabilities) for certain accel and gyro values leading to a crash 2) The Heuristic approach: We specify rules (i.e. fuzzy logic) inferring the accident likelihood from accel and gyro data.
While 1) seems to be the more (mathematically) sound method, gathering empirical data is a lengthy and often infeasible process (crash test dummies anyone?)
Dempster-Shafer's theory of evidence fits into the 1) category having the added benefit of explicitly modelling the 'I-don't-know' case
Crash scenarios that can more or less easily be replicated: 1) Stationary ego vehicle, impact with moving object (car) in 45 degree increments and varying speeds. Special care should be take simulating rear-enders and front impacts
Harder to replicate (but important) scenarios: 1) Moving ego vehicle, impact with moving object; impact angle in 45 degree increments. 2) Accidents without third party involvement, i.e. highsiding, lowsiding These accidents will be hard to replicate but we should see it in perspective of the hurt report statistics:
See road-tests.org for details
What are the crash scenarios we're most likely to encounter and what are the priorities? We should not only look at crash statistics but also at 'ease of implementation' and the likelihood to trigger false positives.