How do you implement a machine learning model to distinguish between pedestrians and static objects in real-time?
Instruction: Discuss the choice of model, the training data required, and how you would ensure the model's accuracy in various lighting conditions.
Updated
Example Answer
I would model object identity and motion state separately. A pedestrian can stand still, and a moving object need not be a pedestrian. Use an appropriate detector to estimate classes and geometry, then associate observations over time to estimate motion with uncertainty. Train and evaluate with labelled edge cases, including stationary people, occlusions and unusual objects. New logged errors can support offline improvements, but the deployed model should not silently retrain itself during operation.
Make it your own
Illustrative labelled cases: a stationary pedestrian, a moving bin, a partly occluded person and a pedestrian about to enter the road. Check class recall separately from motion estimation.
Why this works
This avoids a dangerous binary shortcut from motion to semantics. Follow-up: how would you handle an object that is detected confidently but classified ambiguously?
Sources: TorchVision detection models, OpenCV optical flow, NHTSA ADS safety guidance.