How predict vehicles’ turn direction at intersection?
Predicting Vehicle Turn Intent at Intersections
At an intersection, you are provided with real-time and historical time-series data for N vehicles that are either stopped or moving slowly.
Available Data Per Vehicle:
- Position:
(x_t, y_t)coordinates over time. - Velocity:
(v_{x,t}, v_{y,t})components over time, or alternatively, speed and heading. - Optional Context: Lane and map context specific to the intersection.
Task:
Design a machine learning system to predict each vehicle's intended maneuver at the intersection. The prediction should categorize the intent into one of three classes:
- Left turn
- Go straight
- Right turn
Your explanation should cover the following aspects:
- Label Definition & Ambiguity: How would you define the ground truth labels for training, and how would you address ambiguous scenarios, such as last-second lane changes or indecisive drivers?
- Feature Engineering & Model Selection: Discuss potential features derived from the provided data and suitable model classes, ranging from simple baselines to more advanced sequence models.
- Performance Evaluation: Describe the metrics you would use, strategies for slicing performance analysis, calibration considerations, and methods to prevent data leakage during evaluation.
- Edge Cases & Failure Modes: Identify critical edge cases and potential failure modes of your proposed system.
AI interview prep
Core dilemma
Predicting future intent from noisy, incomplete, and dynamic time-series data, where the 'ground truth' intent can change or be ambiguous until the very last moment, while needing high accuracy and low latency for safety-critical autonomous driving applications.
3 tradeoffs and 3 stress tests locked
Interview prep
Unlock the interviewer X-Ray rubric, tradeoff matrix, stress tests, clarifying questions, and key concepts.
How predict vehicles’ turn direction at intersection?
Predicting Vehicle Turn Intent at Intersections
At an intersection, you are provided with real-time and historical time-series data for N vehicles that are either stopped or moving slowly.
Available Data Per Vehicle:
- Position:
(x_t, y_t)coordinates over time. - Velocity:
(v_{x,t}, v_{y,t})components over time, or alternatively, speed and heading. - Optional Context: Lane and map context specific to the intersection.
Task:
Design a machine learning system to predict each vehicle's intended maneuver at the intersection. The prediction should categorize the intent into one of three classes:
- Left turn
- Go straight
- Right turn
Your explanation should cover the following aspects:
- Label Definition & Ambiguity: How would you define the ground truth labels for training, and how would you address ambiguous scenarios, such as last-second lane changes or indecisive drivers?
- Feature Engineering & Model Selection: Discuss potential features derived from the provided data and suitable model classes, ranging from simple baselines to more advanced sequence models.
- Performance Evaluation: Describe the metrics you would use, strategies for slicing performance analysis, calibration considerations, and methods to prevent data leakage during evaluation.
- Edge Cases & Failure Modes: Identify critical edge cases and potential failure modes of your proposed system.