Sampling the Scenario Feature for Pedestrians Trajectory Prediction

The pedestrian trajectory prediction is an important and necessary task for transportation and vehicle engineering. First, the pedestrians have intention uncertainty and motion freedom. Then, considering the industry requirements, applied methods need to be time-sensitive and stable. The aforementioned properties make pedestrian prediction a great challenge. However, most of current approaches utilize redundant original scenario image in an excessive pursuit of understanding scenario feature. This approach not only diminishes the model's generalization capacity but also amplifies the computing source requirement. In this paper, we propose a novel solution for pedestrian prediction. The various features of scenarios can be embedded with sampling method from planning view like lattice planning [1]. This sampling method utilizes prior knowledge to model valuable scenario information. Moreover, the extraction of local information through sampling can significantly alleviate the computational burden associated with global scenario feature processing in conventional operations. Compared to other methods, our approach maintains computational performance while utilizing fewer parameters and consuming fewer computational resources.