Prediction Flashcards
How do we think about handling multi-modal uncertainty?
Maintaining some beliefs about how probable each potential mode is.
How are these multi-modal predictions represented?
Represented by a set of possible trajectories such as dotted lines and an associated probability for each trajectory.
Name 2 type of Prediction Technologies
- Model Based
2. Data Driven
Explain the Model Based Prediction Technique
Use Mathematical Models of Motion to predict trajectories .
Explain the Data Driven Based Prediction Technique
Rely on machine learning and examples to learn from.
Is Trajectory Clustering a Model Based or Data-Driven Prediction Technique?
Data Driven Approach
Are Process Models a Model Based or Data-Driven Prediction Technique?
Model Based Approach
Explain what a process model is?
This is a Model Based Approach.
A process model is a mathematical description of object motion for behavior.
Are Multi-Modal Estimators a Model Based or Data-Driven Prediction Technique?
Both.
What are Multi-Modal Estimators?
An effective technique for handling the uncertainty associated with prediction, namely, the uncertainty about which maneuver an object will do in a particular situation.
Explain the Hybrid Approaches.
Use data and process models to predict motion through a cycle of intent classification where we try to figure out what a driver wants to do. Trajectory Generation tries to figure out how they are likely to do it.
A prediction module uses what to generate predictions for what all other dynamic objects in view are likely to do?
A map and data from sensor fusion.
Which model is best for determining maximum safe turning speed on a wet road.
In this situation we could use a model based approach to incorporate our knowledge of physics (friction, forces, etc…) to figure out exactly (or almost exactly) when a vehicle would begin to skid on a wet road.
Which model is better at predicting the behavior of an unidentified object sitting on the road.
Data Driven. Even with data driven approaches this would still be a very hard problem but since we don’t even know what this object is, a model based approach to prediction would be nearly impossible.
Which model is better at predicting the behavior of a vehicle on a two lane highway in light traffic.
Hybrid Approach
Trajectory Clustering has two phases.
Offline and Online
Define the offline phase for trajectory clustering.
This is where the algorithm learns a model from data.
Define the online phase for trajectory clustering.
This is where it uses that model to generate predictions.
Name 5 Steps for Target Clustering Offline.
- Get a lot of trajectories.
- Clean the data.
- Define some mathematical measurement of similarity.
- Perform Unsupervised Clustering.
- Define prototype trajectories for each cluster
Name 3 Steps for Target Clustering Online.
For every update cycle:
- Observe vehicle’s partial trajectory.
- Compare to prototype trajectories for each cluster.
- Predict a trajectory.
1 & 2 comparison is done using the same similarity measurement used for offline clustering.
What are Frenet Coordinates?
It is a way representing position on a road in a more intuitive way than the traditional x, y Cartesian Coordinates.
Two main variables: s & d
Explain s in Frenet Coordinates.
s is the distance ALONG the road. Also known as the longitudinal displacement.
Explain d in Frenet Coordinates.
d represents side to side position on the road. Also known as the lateral displacement.
Data-Driven Approaches solve the prediction problem in 2 phases.
- Offline Training
2. Online Prediction
List 3 Steps for Autonomous Multi Modal Algorithm. (AMM)
- Modal conditioned filtering.
- Model Probability update.
- Estimate fusion.
With respect to AMM, modal conditioned filtering does what. (Based on the comparative study)
Runs Kalman filter for each model (m) with initial condition.
With respect to AMM, the modal probability update does what. (Based on the comparative study)
Evaluates the posterior probability for each model (m).
With respect to AMM, the estimate fusion does what. (Based on the comparative study)
Evaluates the overall output - the estimate and the covariance.
How many target motion models are used for each multi modal tracking algorithm.
- 1 non-maneuver model. 8 maneuverable models.
Name the 4 discussed in class.
- Constant Velocity Model - Linear Point Model
- Non-Linear Point Model - Constant Acceleration w/ Curvature
- Kinematic Bicycle Model with Controller - PID Controller with distance and angle
- Dynamic Bicycle Model with Controller - PID Controller on distance and angle
What variables make up the AMM algorithm?
- Consider some set of M process models / behaviors
2. Probabilities for process models (defined as mu)
