Automatic driving multi task awareness network based on improved YOLOP

In recent years, deep learning technology has developed rapidly and is widely used in the field of autonomous vehicle driving. Multi-task networks can simultaneously complete multiple tasks in autonomous driving environment perception, and can minimize hardware costs while being highly efficient, so they are very popular in low-cost autonomous driving systems. This paper improves the problems existing in the YOLOP network and proposes an efficient multi-task autonomous driving perception network, which uses a shared encoder and three independent decoding heads to simultaneously complete road traffic target detection, drivable area detection and lane detection. Three major perception tasks of line detection. We use the SPPF module in the backbone network instead of the SPP module in YOLOP, so that the network can obtain feature information of different scales while being lightweight; secondly, in the Neck layer, we abandoned the classic structure of FPN+PANet in YOLOP and proposed the HDM structure Rich receptive field information can be extracted, and multi-scale feature fusion tasks can be cleverly completed by cascading with shallow networks. Finally, in the experiment, we found that different tasks of the multi-task network have different requirements for feature information. Each task We cannot simply use the deepest features of Neck like YOLOP, so we adjusted the network structure and input the required feature information for different tasks. The final results show that the multi-task network we designed can achieve an astonishing 97 FPS on NVIDIA TESLA V100, which is far more than the 49 FPS of the YOLOP model under the same experimental settings, ensuring that autonomous vehicles can quickly and accurately perceive the surrounding environment., which is conducive to the safe and reliable driving of autonomous vehicles.