ABSTRACT

Accurately predicting the crossing behaviour of pedestrians remains a significant challenge due to their complex behavioural dynamics. Although modern transformer-based models have shown promise in being able to accurately capture these dynamics, the crucial role of contextual information, especially under occluded scenarios, has been underexplored. In this work, we demonstrate that additional contextual features, such as crosswalk visibility and traffic light status, can assist in improving prediction performance under degraded conditions, where accurate pedestrian information is not readily available. We propose AugTrEP, inspired by the existing Transformer-based Evidential Prediction (TrEP) network, which uses two transformer encoders with cross-attention to learn pedestrian behaviour by incorporating global traffic context. We evaluate our models against the PIE benchmark and curated test sets simulating the behaviour of real-world perception systems under varying degrees of occlusion. Our analysis reveals a significant improvement in the accuracy, AUC, F1 score, and precision compared to the baseline under degraded input conditions. These findings highlight AugTrEP’s resiliency to disturbances caused by occlusions and emphasize the importance of scene context in accurate behaviour prediction for real-world applicability.

Month: May

Year: 2024

Venue: 21st Conference on Robots and Vision

URL: https://crv.pubpub.org/pub/ed7j1xf9