Sidewalk-level People Flow Estimation Using Dashboard Cameras Based on Deep Learning

Abstract

In future smart cities, understanding people flow is essential for a wide variety of applications such as urban planning and event detection. In this paper, we propose a method to estimate low of pedestrians walking on the sidewalks. Our goal is to achieve high accuracy with wide area coverage at low cost. For this purpose, we use dashboard cameras (dashcams) recently wide-spreading for providing evidence on accidents, etc.Our method combines Deep Learning-based pedestrian detection and model-based tracking to overcome the challenges of frequent occlusion (overlaps of objects in images) and false positives. In pedestrian detection, faces and backs of heads are separately detected to understand moving directions of pedestrians as well as their existence by applying CNN (Convolutional Neural Networks) and LSTM (Long-Short-Term-Memory). Then, the trajectories of the detected bounding boxes are estimated based on location and color similarities with the knowledge about moving speeds of vehicles and pedestrians.Through the evaluation using real dashcam videos, we have confirmed that our method works well, showing within ±13.6% mean absolute error rate of bi-directional people low estimation.