search

dvij542 / trajectory-prediction-GRIP

Improve upon GRIP++
7 stars 2 forks source link

Explicit maneuver inference #3

Open dvij542 opened 3 years ago

dvij542 commented 3 years ago

After the encoder layer, the output (NV, C) is inputted to find the lateral and longitudinal maneuver, refer to the CSP code for this. The (NV,C) must pass through separate decoder layers to get 6 trajectory outputs for each each instance. Backpropagate through the ground truth maneuver class (Refer to the CSP code) and during validation, use the class predicted by the explicit class prediction layer.

CSP code : (https://github.com/nachiket92/conv-social-pooling)

dvij542 commented 3 years ago

NGSIM preprocessed dataset can be downloaded from (https://drive.google.com/file/d/1K8HosrYXvKN7EBdsqogUiGXuGUfw5MRM/view?usp=sharing) with added lane data. The format can be seen from comments in preprocess_ngsim.py file.