Spatiotemporal Trajectory Prediction for Traffic Agents

About

This project implements several machine learning algorithms relative to traffic agent trajectory prediction.

The most recent model implemented is a Seq2Seq model with a PointNet model used to encode the environment. Though simple, it gives competitive results on the ArgoVerse dataset.

Task List

Data Engineering:

• Agent-Centered Transformations
• Orientation from start/end of input sequence
• Orientation from displacements
• Lane Relevancy Metrics (Rear, Orientation Alignment, Distance, etc.)

Backbones:

• MLP
• RNN
• Seq2Seq (LSTM, GRU, etc.)
• Transformer
• Graph Neural Network

Spatial Encodings:

• Sinusoidal Positional Embeddings (MLP)
• PointNet
• PointNet++
• Graph Convolutional Network
• ResNet
• CNN
• ConvLSTM

Loss Functions:

• ADE/FDE Loss

Other Approaches:

• Social Pooling
• SocialGAN
• STGAT
• Trajectron
• TPNet

Contact

Yishai Silver (ssilver@ucsd.edu)