Pedestrian Trajectory Prediction using BiLSTM with Spatial-Temporal Attention and Sparse Motion Fields

Autonomous cars and mobile robots need to be able to anticipate and interpret pedestrian intents in order to manoeuvre safely in a congested area. However, because pedestrian movement is unpredictable, it is necessary to take into account a variety of elements when modelling their potential trajectories, including their prior movements, interactions with other pedestrians, and restrictions imposed by static objects in the environment. Many current trajectory prediction techniques ignore the possible influence of static impediments on pedestrian movement, and instead have concentrated only on how pedestrians interact with one another in a given scenario. Furthermore, many of these techniques need additional data such as semantic maps, which reduces scalability and may not always be available. Our proposed model, named scene and social interaction using biLSTM, Spatial-Temporal Attention and Sparse Motion Field (SS-BLSTAS), uses simply the trajectory of pedestrians as input and is accurate at predicting future movement, by considering both the presence of obstacles and social interactions. The proposed model is compared to other state-of-the-art models by using standard evaluation metrics on the ETH and UCY pedestrian datasets. The experimental results show that the proposed model performs better than most of the benchmarked approaches.