Combining LSTM and MDN Networks for traffic forecasting using the Argoverse Dataset

Trajectory forecasting is vital to target tracking, autonomous decision making, and other fields critical to the future of autonomous systems. Tracking algorithms, such as the Kalman Filter, require accurate motion models in order to forecast target trajectories and update state estimates given observation data. Unfortunately, accurate motion models are not always easily de- fined. Of particular interest is forecasting in systems with complex agent-to-agent and agent-to-scene interactions, which are often best represented as a multimodal distribution. Various network architectures tackle this multimodal problem in different ways, but the method used in this work is a mixture density network. The network architecture examined in this work, LSTM2MDN, builds off previous research in combining the renowned long- short term memory (LSTM) network with a mixture density network (MDN) in order to develop accurate distributions for output trajectories.