Region-based Deformable Fully Convolutional Networks for Multi-class Object Detection at Signalized Traffic Intersections

Multi-class object detection is critical for intelligent traffic monitoring applications in smart cities as well as connected autonomous vehicles. Although, numerous research works evaluate the performance of image processing algorithms for onvehicle cameras, the body of research evaluating performance of image processing of stationary cameras located near intersections is limited. In this research, we use region-based deformable fully convolutional networks to detect 14 different object classes within images from traffic surveillance cameras. The object classes include vehicles, pedestrians, bicyclists and traffic signals. The goal of the NVIDIA AICity challenge is to provide accurate localization and correct classification of objects for stationary cameras mounted near signalized traffic intersections. Our proposed method scores mean average precision of 0.41, 0.37, and 0.34 for aic480, aic1080, and aic540 challenge datasets. Keywords—multi-class object detection, surveillance camera, deformable convolutional networks, and deep learning