|Driving speed as a hidden factor behind distribution shift|
|Organisatsiooni nimi||AUtonomous Driving Lab|
|Kokkuvõte||When we drive at different speeds, the inputs and outputs may differ:|
- If we consider consecutive frames as input, the difference between these frames is bigger when driving faster
- If we consider “change image”/optical flow as input, the differences are amplified at higher speeds
- Steering angle distribution can be different at different speeds. This might be due to a lag of commands taking effect (at higher speeds you need to send turn command earlier).
In the simplest AI-driving solutions we collect data at fixed driving speed X and train the neural network to imitate human steering on the collected data, based on camera image (steer=f(image)). We then test if this steering predictor actually can drive at speeds X, 2X and 0.5X.
We hypothesize that the input data and the labels collected at speed X might not allow the model to drive at 0.5X nor 2X. At those speeds the model would:
1) encounter input data it has never seen before (e.g. difference images with more difference)
2) would predict commands too early or too late as it attempts to counter lag at 1X not at the current speed.
In the real-life, with the ADL Lexus we collect data at 60km/h+ but sometimes test the models at 30km/h for safety. Does using slower speed actually hurt a model’s ability to drive?
We can test the effect of speed difference on the toy cars, plot and analyze it thoroughly. There are existing metrics for detecting if inputs are “out-of-distribution” for a network. Once the effect is proven on toy car data, to make this publishable, we can proceed to demonstrate the same on the real car.
|Lõputöö kaitsmise aasta||2021-2022|
|Suhtlemiskeel(ed)||eesti keel, inglise keel|
|Märksõnad||#deep learning, self-driving,|