Exploration And Research On Application Of Deep Neural Networks In Automatic Driving Of ROS Based Self-built Unmanned Vehicles

The huge amount of data generated by self-driving system and the amount of computational consumption demand high performance from hardware platform.For example,current researches on object detection mainly focus on improving the speed and accuracy of the algorithm,regardless of power consumption,as well as costs.This thesis proposes a low-cost solution to build an unmanned vehicle platform for self-driving researches based on the Robot Operating System.Based on this platform,we try to research on how to improve the speed of mainstream object detection algorithm and end to end self-driving in indoor environment.This thesis introduces the depthwise separable convolution to improve the single shot detector algorithm,reducing the amount of parameters in its convolutional layers,thereby improving the speed of the network on the low-power platform and reducing the prediction delay.Secondly,this thesis discusses the possibility of discarding the largest feature map of multiple feature maps in the SSD algorithm,in order to sacrifice the detection performance of some tiny objects in exchange for the detection speed on low-power platforms.The experimental results show that the prediction delay of the improved SSD model under this platform is reduced by about two-thirds,with only an average drop of 10% on detection confidence.This thesis also discusses the research of end-to-end automatic driving in indoor environment on the platform.Based on the relevant work of Nvidia,this thesis introduces a new lightweight model with dilated convolution and depthwise separable convolution.This model improves the perception towards the indoor environment with extended receptive fields,meeting the requirement of low-power consumption.The experimental results show that the improved model can better identify and separate highly repeated objects with each other in indoor environment.The overall control rate of continuous cruise is about 97.8%,achieving an 8.7% improvement compared with that of Nvidia's initial model.Training time also decreases due to faster convergences.