Learning to Detect Pedestrian Flow in Traffic Intersections from Synthetic Data

Detecting the pedestrian flow from different directions at a traffic intersection has always been a challenging task. Challenges include different crowd densities, occlusions, lack of available data and so on. Emergence of deep learning and computer vision methods has shown potentials to deal with this problem. Most of the recent works focus on detecting combined pedestrian flow or counting the total number of pedestrians. In this work, we propose to detect not only combined pedestrian flow but also pedestrian flows from different directions on a crosswalk at a traffic intersection. Our main contributions are summarized as follows: 1) we introduce a synthetic dataset that we create using computer game and a real-world dataset we collect from the street; 2) we propose a Pedestrian Counting Network (PCNet) to count pedestrians from different directions along a crosswalk and then propose a Pedestrian Flow Inference Model (PFIM) to infer the pedestrian flow parameters of volume and density; 3) we design a structure-aware domain adaptation for learning from synthetic data to real data. The experimental results show the effectiveness and accuracy of the proposed method.