|Organisatsiooni nimi||Autonomous Driving Lab|
|Kokkuvõte||End-to-end driving network needs to take into account other cars, pedestrians, traffic signs, traffic lights, etc. Learning how to react to all those objects with just the supervision signal from driving commands is challenging. Therefore it makes sense to provide the additional supervision signal to the network by making it to predict cars, pedestrians, traffic signs, traffic lights, etc. in addition to the driving commands (usually steering wheel position and speed). This can be done by making the neural network multi-head and predicting multiple outputs at once. This forces the network to represent various road objects, which hopefully is useful for the final driving task. The training targets for additional outputs can be taken from other specialized networks, for example object detection and segmentation networks.|
The project consists of:
1. Literature review of multi-head networks.
2. Investigation of different object detection and semantic segmentation networks to include for target prediction.
3. Collecting and preparing a dataset of driving videos.
4. Training a neural network to predict multiple outputs from a single image.
5. Training an end-to-end driving neural network on top of the representations of the multi-head network.
|Lõputöö kaitsmise aasta||2020-2021|
|Suhtlemiskeel(ed)||eesti keel, inglise keel|