Research On Free Space Detection Method For Intelligent Vehicle Based On Deep Learning

Intellectualization is the development direction of modern automobile technology,and environmental perception is the basis and key technology to intelligent vehicle.The free space detection of road is the main content of environment perception research,and it determines the automation level of intelligent vehicles.With the rise of deep learning,the remarkable achievements of deep convolutional neural network in target recognition have directly affected and promoted the development of environmental perception technology for intelligent vehicle.However,because the performance of the network is restricted by its space and time complexity,this contradiction makes it difficult for the existing deep network to meet the requirements of environmental perception system for both accuracy and speed when applied to intelligent vehicle.Therefore,under the current theoretical research depth of deep convolution network and the development level of on-board sensor technology,it is still a challenging task to realize a robust,accurate and fast road free space detection algorithm based on deep learning.Under the above background,this paper relies on the major science and technology platform project of Liaoning Provincial Department of Education,"Research on Environment Perception Technology Based on Vehicle Radar and Vision Sensor Information Fusion"(JP2016018),focusing on the "person-vehicle-road" environment system,and based on deep learning method,the key technologies in free space detection: road detection,pedestrian detection and multi-target joint detection are studied.The specific research work of this paper includes:Firstly,road detection is the core content of free space detection,and fast and accurate road detection results are the basis of intelligent vehicle trajectory planning and the key to ensure safe driving of vehicles.Aiming at the problems of high network complexity and low computational efficiency caused by the adoption of pyramid structure and attention mechanism with multi-level recursion in existing road segmentation methods based on deep learning,a road segmentation method based on spatial attention transfer mechanism is proposed.According to the distribution law of features between convolutional neural network layers,this method combines the non-local spatial attention mechanism with the channel attention mechanism,and adopts add-and-multiply strategy to transfer the long-range dependencies from the bottom features to the top features,which effectively improves the receptive field of features and reduces the computational complexity.At the same time,a 1×1 convolution kernel dropout method is designed to avoid network over-fitting on the premise of making effective use of attention mechanism.The test results on Cityscapes data set show that the proposed method can significantly improve the detection speed while maintaining high detection accuracy compared with the existing similar methods,and have better generalization ability in cross-data set test.Secondly,pedestrian detection is an important part of free space detection.Real-time and accurate pedestrian detection results are the basis for intelligent vehicles to make behavior decisions and the premise to avoid personal injury.Aiming at the problem that the cascaded multi-feature fusion structure adopted by the existing deep learning-based pedestrian detection methods cannot give full play to their respective advantages,resulting in poor robustness to target occlusion and apparent state changes,a pedestrian detection method based on cascaded adaptive boosting algorithm and deep convolution network is proposed.In this method,a cascade AdaBoost classifier is designed as a region proposal method and combined with CNN to form a two-stage target detection network.A fast aggregation channel feature pyramid is adopted to obtain multi-scale features efficiently,and a negative sample retrieval strategy is designed by using the feature of cascade AdaBoost detecting positive samples step by step.Difficult samples are randomly selected according to a certain proportion and sent to CNN for classification,thus realizing the complementary advantages of efficiency and accuracy between cascade AdaBoost and CNN.The test results on Caltech data set show that compared with similar methods,the missing detection rate of this method is significantly reduced while maintaining a higher detection speed.The cross-dataset test on the actually collected campus driving video has good robustness to the change of pedestrians.Finally,on the basis of the above research,a multi-task detection algorithm based on spatial correlation feature propagation is proposed to solve the problem that the real-time performance is difficult to be further improved due to a large number of redundant calculations in the current multi-task method of video processing.Based on the similarity of inter-frame images in video,a feature propagation method guided by spatial correlation is designed.The offset representing spatial change is obtained by Sobel operator,and the high-level features of key frames are obtained by differential bilinear interpolation,which effectively improves the overall computational efficiency of feature extraction.Based on this feature propagation method,a multi-task network framework for road segmentation and vehicle and pedestrian detection is constructed,and the parameter training of feature propagation network is enhanced by using the high-level feature consistency constraint design auxiliary training network,which improves the modeling ability of feature propagation network for feature changes.Tests on representative KITTI and Cam Vid datasets show that the method proposed in this paper has achieved the same accuracy as state of the art algorithm in road segmentation and pedestrian and vehicle detection,but the speed has been significantly improved.It has better robustness to the changes of road shadows and pedestrians in the cross-dataset tests on the actual urban road videos.