Application Of Lightweight Convolutional Neural Network In Intelligent Driving Awareness System

Intelligent driving is the trend of automobile development,and it is also a hot spot in the current automotive industry research.Environmental awareness is an important module in the intelligent driving system.Its main task is to perceive pedestrians,vehicles,traffic lights,obstacles and other objects around the intelligent driving car,so that the intelligent driving system can make correct path planning to avoid collisions when the car is driving.Intelligent driving vehicles have high requirements on the real-time performance of the environment-aware module,while some current mainstream models aim to improve the accuracy,ignoring the real-time nature of the model.Lightweight technology is an important technical means to optimize the real-time performance of the model.This paper is mainly based light-weight technology in deep learning,research in environmentally-aware of the image perception,proposed a set of visual recognition and semantic segmentation scheme for smart driving scene.First,for the problem of visual recognition,this paper based on light-weight and efficient convolution methods such as deep separation convolution and group convolution,design invariant resolution convolution module and a down-sampling module for image feature extraction,and then construct the backbone network.For the detection part of the network,we have made improvements and optimizations on the SSD network.And trained,tested and analyzed the hole network on the KITTI dataset.Secondly,for the semantic segmentation problem,this paper designs a deep convolutional neural network for semantic segmentation using efficient multi-scale information fusion unit and efficient lightweight network.And then trained and tested the model on the CityScape dataset.Lightweight techniques such as pruning and quantification compress the model to optimize network speed and model size.Finally,the visual recognition model was experimentally verified on the KITTI dataset,and 72.7% of mAP were obtained.The inference speed reached 66.7 FPS on the NVIDIA 1080 Ti GPU.The semantic segmentation model was experimentally verified on the CityScape dataset.The semantic segmentation model mIoU indicator is between 65~70%,the inference speed can reach 20 FPS on NVIDIA 1080 Ti GPU,and the inference speed can reach 38 FPS on NVIDIA Tesla GPU.