Pedestrian Crossing Intention Prediction Using Transformer-Based Feature Fusion

Pedestrian intention prediction can be used in Advanced Driver Assistance Systems to prevent pedestrian-vehicle collision in case of driver distractions. The use of these tools will reduce pedestrian fatalities in traffic accidents. In addition, predicting pedestrians' intention and avoiding collisions are parts of the functions of autonomous vehicles. Improving these models will lead to safer vehicles and higher levels of automation. In most studies, inspired from human behavior, sequences of spatial features, e.g. pedestrian's location, are considered as inputs. For more accurate prediction, temporal features are also extracted from these sequences of spatial features. We have realized that it is important to represent key objects to ensure model's perception of the scene. In this work, a spatio-temporal approach has been studied that predicts pedestrians' intention using spatial features of cropped images. We use the same strategy as our reference model for temporal features and feature fusion. We have modified the masking model of the reference model to reduce inference time, while we improve the performance. With inference time of 58 ms, we achieved accuracy, AUC and F1-Score of 0.90, 0.87, and 0.75 on JAAD dataset, respectively.