Deep learning approach to predict vehicle movement of self driving cars
This was my final project for CS7643 Deep Learning. I worked in a group of four of my friends and we really had a blast with this. Basically, we implemented a paper which used lots of deep learning techniques to predict the motion of vehicles. It taught us about multi headed attention, LSTMs, and a concept known as social tensors. Social tensors are basically a way to model interactions between agents in space, originally used to predict how people would move throughout a room, but in our case is applied to the interaction between cars. The general idea of this algorithm was to combine social tensors with a feature representation of the birds eye view to predict future movement. Picture 2 shows the target vehicle in green, other vehicles in blue, and the prediction for the target vehicle in purple. Ultimately, we did not achieve great accuracy, likely because we still have bugs (even after countless nights of debugging). Nonetheless, this was certainly the most heavy deep learning project I've done, and it considerably improved my pytorch and paper reading skills!