Pedestrian Cross Forecasting with Hybrid Feature FusionDownload PDF

Published: 07 Apr 2023, Last Modified: 14 Apr 2023ICLR 2023 Workshop SR4AD HYBRIDReaders: Everyone
TL;DR: multi-modal for pedestrian crossing intention prediciton
Abstract: Forecasting the crossing intention of pedestrians is critical for the growth of Autonomous Vehicles (AVs) in the real world. Pedestrians' movements are usually influenced by their surroundings in traffic scenes. Recent works extract explicit and individual information from collected data to perform prediction. However, there still exists much implicit information which is not considered ever, such as interactions between features, location of pedestrians, and distance towards the ego-car. Properly exploring and utilizing the implicit information will promote the prediction of future behaviors. To this end, the surrounding interactions from semantic segmentation and local context, together with two novel introduced attributes: the pedestrian's location at road or sidewalk, the relative distance from target pedestrian to ego-car are adopted as critical features in this paper. The location and distance attributes are derived from the semantic map and depth map combined with bounding boxes information separately. A hybrid network based on multi-modal, which incorporates interactions between individual features, is proposed to forecast cross or not. Evaluated by two public pedestrian crossing datasets, PIE and JAAD, the proposed features and fusing strategy achieve state-of-the-art performance.
Track: Research Insight
Type: PDF
0 Replies

Loading