Dense-ACSSD for End-to-end Traffic Scenes Recognition

Traffic scenes recognition is the cornerstone of autonomous driving. However, most of the current algorithms are individually trained for tasks such as object detection and road segmentation. In addition, the training data used are mainly concentrated in small datasets such as KITTI, and the trained models are highly susceptible to weather, lighting and other factors. In order to solve the above problems, we propose an end-to-end CNN model for drivable area segmentation and multiple object detection. The feature extraction part of the network is powerful DenseNet. The atrous convolution and spatial pyramid pooling are used for road segmentation, and single shot detection is used for multiple object detection. According to its characteristics, we named the network Dense-ACSSD. Dense-ACSSD is trained on the current largest autonomous driving dataset, called BDD100K. The final training results show that the mIOU of the drivable area segmentation part is as high as 84.15%, and the mAP of the multiple object detection part reaches 30.82%. In addition, the inference time of Dense-ACSSD can meet real-time requirements.