Research On Classification Detection Algorithms Of Traffic Objects In Road Driving Scenes Based On Heterogeneous Vision

With the rapid development of autonomous driving and advanced assisted driving technologies,the detection of traffic objects in road driving scenes based on visual environment is indispensable.It is particularly important to improve the intelligence and safety of intelligent transportation and to achieve safe,reliable and comfortable driving tasks.Extensive and sufficient perception of vehicle surrounding environment information,obtaining detection information of road traffic objects,improving road travel safety,and reducing traffic safety accidents are important contents of current research in the field of intelligent transportation,and it has become a hot topic of research and attention by scholars in this field.Based on a full investigation of the current research status at home and abroad,this paper surrounds the needs of road vehicle environment perception and many problems in the detection of traffic objects in road driving scenes.It focuses on the classification detection methods of the three elements of road traffic(road,vehicle,and pedestrian)under different visual sources.The main research work and contributes can be summarized as follows:In order to solve multiple problems that most road detection methods are sensitive to illumination changes and shadows,and the traditional illumination-invariant road detection methods have problems such as the difficulty of determining the camera axis calibration angle and the distortion of road sample sampling,an online detection algorithm for shadowed roads using learning-based illumination-independent image is proposed.First,the road blocks and non-road blocks in the road image sequence are demarcated manually,and the multi-feature fusion method is used to train and generate the SVM classifier for road blocks.Then,the RGB space of the combined road block is converted into the geometric mean logarithmic chromaticity space,and the camera axis calibration angle of each frame of image is determined according to Shannon entropy,and the illumination-independent image of each frame is obtained.Finally,the road sample points are extracted by the random sampling method of the safety distance area in front of the vehicle,the road confidence interval classifier is established,and the road is separated from the background.Multiple video sequences are used to evaluate the effectiveness of the algorithm.Experimental results show that this method can obtain high quality road detection results and realize the real-time performance of road detection.Aiming at the problems of large matching error of binocular disparity figure in real traffic environment,high computational complexity,and lack of necessary depth information for the detection of traffic objects such as vehicles,vehicle detection on the road algorithm based on monocular depth estimation and considerate U-V view is proposed.The original disparity figure is obtained by using the monocular depth estimation model added with edge enhancement loss function,and the initial U and V views are defined by the horizontal and vertical projections of the original disparity figure.In order to obtain the region of interest(ROI)of vehicles on the road,an algorithm of ROI detection based on considerate U and V views is proposed.For the sake of obtaining the detection results of vehicles on the road,a parallel scanning algorithm for road areas is proposed to determine the source points of vehicles or pedestrians on the road.At the same time,the neighboring disparity similarity algorithm is used to complete and extract the target area,and the extracted target area is accurately segmented by combining multiple feature fusion methods such as aspect ratio,perspective ratio and area ratio to obtain vehicle detection results on the road.The experimental results show that this method can realize the classification detection of vehicles or pedestrians under the same detection framework and meet the time validity of vehicle detection on the road.In view of the ineffectiveness of visible light images in the detection of night traffic objects,especially pedestrian detection,and the contradiction between the accuracy and real-time performance of existing night pedestrian detection methods,a night pedestrian detection algorithm based on rapid saliency and multi-feature fusion is proposed.A saliency model for pedestrian targets in infrared images is used to achieve rapid acquisition of pedestrian target areas at night.A method for refinement and separation of target regions is proposed to remove the possible non-pedestrian area adhesion interference in the region of interest,and obtain the pedestrian candidate bounding box at night.The classification of pedestrian at night is realized by multi-feature fusion feature extraction method and SVM classification algorithm,and use multiple video sequences to evaluate the effectiveness of the night pedestrian classification algorithm.Experimental results show that this method can obtain high quality classification detection results of pedestrians at night and meet the real-time requirements of road scenes.The research work can realize the classification detection of traffic objects(road,vehicle and pedestrian)in road driving scenes based on heterogeneous vision,and provides a new mode of traffic environment perception for road driving scenes.It has a certain significance to help improve the safety of vehicles on road traffic,and provides an effective method for building a road environment perception system based on visual information.