Multi-network for Joint Detection of Dynamic and Static Objects in a Road Scene Captured by an RGB Camera

Abstract

AbstractThis study presents a unified network to localize dynamic, static traffic objects and pedestrians, classify traffic light colors, detect drivable area and lane line simultaneously. In the network architecture, traffic object branch is created to classify dynamic objects such as cars, trucks, buses, motorcycle, and bicycle. Static objects are categorized by traffic sign and traffic light objects. Pedestrians are also localized as a separate traffic object group. Traffic light is classified correctly when it is visible. The network design has a unified architecture, one shared encoder for feature extraction and three decoders for three tasks. For benchmarking purposes, the BDD100K dataset is used. The presented model is ranked in the second place for drivable area segmentation, lane line detection, and inference speed while benchmarking with publicly available multi-networks. In comparison with respect to state-of-the art segmentation models re-trained with BDD100K dataset, the task of dynamic object localization’s MIoU metric is reached to the level of 73.54%, which is 40% higher than the results of re-trained segmentation methods.KeywordsDeep learningMulti-task learningDynamic and static traffic object localizationPedestrian localizationTraffic light classificationDrivable area segmentationLane line detection