Efficient And Robust Convolutional Neural Networks For Autonomous Driving

Autonomous driving has drawn tremendous attention in recent years.Affiliates,such as high-tech companies,automobile companies,and research institutes,from either domestic or abroad have been developing and experimenting autonomous driving techniques.Driving environment is an open environment that has various scenarios,weathers,objects and stuff.Facing such environment of driving,it is essential that the sensing models be capable of handling various scenarios stably in real time.Those results from all sensing models shall be fused with proper strategies for the final driving decision and operation.Despite the great success of convolutional neural networks on computer vision tasks,such models come with high computational costs and large memory consumption.To meet those requirements,expensive high-performance parallel computing devices and reliable power supplies are needed,which further limits the application of deep learning in autonomous driving cars.This paper is mainly aimed at studying on deep learning algorithms for high-level autonomous driving,specifically on convolutional neural networks for visual sensing tasks.A neural network framework which can reduce the uncertainty of sensing models,called Uncertaity Learning Neural Network(ULNN),is proposed in this paper.Uncertainty learning layer and uncertainty loss function are designed in ULNN and can significantly lower the uncertainty of visual sensing models in autonomous driving systems.The proposed ULNN suites intelligent systems which require multi-model fusion as autonomous driving system does.To tackle the drawbacks of a too high computational cost to meet the real-time demand for autonomous driving,efficient convolutional neural network architectures(named GlanceNets)with multiple output bypasses are designed in this paper.The bypasses can early output the classification results with the help of the prediction confidence measurement based on information entropy of Softmax outputs.To make the proposed architecture able to mine hard examples online,we develop hard example weight function inspired by online hard example mining methods and designed an adaptive threshold learning method for all bypasses with one adjustable balancing parameter.The key contributions of this paper are:(1)Several mainstream convolutional neural networks that are suitable for autonomous driving tasks are studied;(2)ULNN,which is capable of reducing the uncertainty of visual sensing models in a learning method,can improve the robustness of sensing results for autonomous driving;(3)The proposed GlanceNets,efficient convolutional neural networks with adaptive hard example mining,can sufficiently reduce the computational cost and can be used in various vision sensing tasks.