Pedestrian Trajectory Prediction Using A Social Pyramid

Understanding and forecasting human movement paths are vital for a wide range of real world applications. It is not an easy task to generate plausible future paths as the scenes and human movement patterns are often very complex. In this paper, we propose a social pyramid based prediction method (SPP), which includes two encoders to capture motion and social information. Specifically, we design a social pyramid map structure for the Social encoder, which can differentiate the influence of other pedestrians in nearby areas or remote areas based on their spatial locations. For the Motion encoder, a mixing attention mechanism is proposed to combine the location coordinates and velocity vectors. The two encoded features are then merged and passed to the decoder which generates future paths of pedestrians. Our extensive experimental results demonstrate competitive prediction performance from our method compared to state-of-art methods.