ShuDA-RFBNet for Real-time Multi-task Traffic Scene Perception

Traffic scene perception is key of autonomous driving. Computer vision and deep learning are popular basis of algorithms in this field. Current researches are mainly on single task, leading to high computing power demands and low operational efficiency. And algorithms may be trained on a small dataset, which means easier overfitting and more susceptible to noises. To overcome above shortcomings, we propose an end-to-end multi-task deep convolutional network model named ShuDA-RFBNet for multiple object detection and drivable area segmentation simultaneously. ShuDA-RFBNet adopts RFBNet for multiple object detection and DenseASPP for drivable area segmentation task. Both networks share a light-weight ShuffleNet V2 as the base net to extract features. ShuDA-RFBNet uses BDD100K to train, which is the current largest and diverse driving video dataset. Test results show that the mAP of multiple object detection task is 32.71% and the MOU of drivable area segmentation task is 82.67%. ShuDA-RFBNet can inference in real time.