MultiPath++: Efficient Information Fusion and Trajectory Aggregation for Behavior Prediction

MPA: MultiPath++ Based Architecture for Motion Prediction

“…For training our model we first did some preprocessing for agents of interest. Since our model is based on Mul-tiPath++ [16] we follow similar process of data input data preparation as in MultiPath++ [16]. A usual approach is to first transform the frame into the canonical coordinate system where the agent we make prediction for is always located in the same position with the same heading at the moment we make prediction for.…”

“…This step helps us to eliminate redundant symmetries. Road graph data was represented as in MultiPath++ [16]. For the target agent and other agents that surround the target agent for each timestamp in history (past + current) we computed x, y coordinates, heading, velocity in the mentioned canonical coordinate system and the validity boolean flag.…”

“…Our model follows the architecture of MultiPath++ [16] mostly, however we tried to alternate it a little (see Fig. 1).…”

“…Second we did not use proposed EM algorithm because in our experiments it performed numerically unstable. Instead we experimented with a single decoder with 6 modes and 5 decoders each with 6 modes followed by attention mechanism [17] and multi-context gating block (MCG) from MultiPath++ [16] that mapped 30 modes into required 6. With multiple decoders we used a proposed strategy of blocking weighs update for randomly selected decoders (see Fig.…”

“…One of the key difficulties is a natural uncertainty of other agents' behavior. In this paper we propose a solution based on MultiPath++ [16] which efficiently scores the 3rd on Waymo Motion Prediction Challenge 2022 and significantly outperforms our previous solution based on Convolutional Neural Network.…”