Design Of Vision Perception Algorithm For Autonomous Driving Based On Monocular Vision

In recent years,with the development of artificial intelligence,autonomous car has become a hot topic.As one of the indispensable sensors in autonomous driving system,cameras provide image information for the autonomous driving system.The autonomous driving system can acquire information such as the position of various obstacles and traffic signs,as well as drivable areas by cameras.This paper focus on object detection task and drivable areas segmentation task based on monocular vision for the autonomous driving system.Deep learning has greatly push forward the development of computer vision.Convolutional neural networks have greatly improved the accuracy of vision-based object detection and semantic segmentation,but the computational complexity of deep convolutional neural networks is enormous.Since the autonomous driving system needs to control the car in real time,this requires the visual perception algorithm to achieve real-time inference,and the autonomous driving system run on mobile platform,and the computing power of mobile platform is very limited.Therefore,this paper designs a multi-task convolutional neural network to detect objects and segment the drivable area.The main research contents of this paper as follows:An image classification model Net-A is designed as the backbone network of the object detection network and the drivable area segmentation network.In this paper,we analyze many state-of-art image classification models,and analyze the computational performance of the autonomous driving platform.We design a image classification model Net-A,And train Net-A on the ImageNet dataset.The Top5 classification accuracy rate of Net-A is lower than ResNet-50,but is higher than VGG-19.The inference speed of Net-A is higher than ResNet-50 and VGG-19.We design an object detection network and a drivable area segmentation network based on Net-A.In this paper,we optimize the architecture of YOLOv3,and improve the inference speed.Based on the idea of FCN model,we design the drivable area segmentation network.These two networks are trained on the BDD100 K dataset.The inference speed of these two networks at 720?1280 resolutions are 46.7 FPS and 49.2 FPS,respectively,far exceeding the real-time requirements for autonomous driving system.The performance of object detection network surpasses YOLOv3-418,and the segmentation accuracy and inference speed of the drivable area segmentation network are faster than DRN-D-22 and ERFNet.We combine the object detection network and the drivable area segmentation network.Running the object detection network and the drivable area segmentation network in a same time will consume a large amount of computing resources.Considering that both the object detection network and the drivable area segmentation network are based on the same backbone network Net-A,we combine these two networks by sharing the parameters of the backbone network,and obtaining a joint model of the object detection network and the drivable area segmentation network.Although the joint model accuracy is slightly lower than the single model of the object detection network and the drivable area segmentation network,but the inference speed of joint model reaches 37.5FPS.We deploy the object detection and the drivable area segmentation joint network on the NVIDIA GTX1070 GPU computing platform of which computing performance is weak.Due to the weak computing performance of the NVIDIA GTX1070 GPU,the joint model can’t run in real-time.Therefore,we optimize the computational structure of the joint model,reducing redundant computational nodes of the model.Finally,the inference speed of joint model can reach 23.21 FPS with an input resolution of 720?1280 on the NVIDIA GTX1070 GPU computing platform.