Research On Key Technologies Of Vehicle Intelligent Driving Based On Visual Perception

As one of the key technologies in the field of intelligent driving,perception technology determines the ability and intelligent level of autonomous driving of automobiles.Pedestrian detection is an important part of the perceptual detection task.Owing to the low cost and good performance of the camera,pedestrian detection technology based on visual image has been widely studied.Therefore,the difficulty of pedestrian detection based on visual image is studied.The main contents of our research are as follows:The development of general target detection and pedestrian detection models are compared and analyzed,and the basic theories of target detection based on deep learning are studied,including general network architecture and training strategies,as well as autonomous driving dataset.In order to solve the problem that small targets in large images are easily missed,a road pedestrian small target detection method based on image partition is proposed.Firstly,combined with City Persons dataset characteristics,a new pedestrian category label is designed,and a partition dataset is constructed.Furthermore,the results of direct pedestrian detection by YOLO v3 and v4 are compared and analyzed.Secondly,based on YOLO v4,the strategy of image partition detection and merging is completed.When the cropping ratio is 0 under the image partition detection,the AP value of the small target pedestrian SR＿person is increased from 35.29% for direct detection to61.15%.Finally,in order to improve the ability of pedestrian feature representation,the model uses the USC dataset pre-training,the AP value of the small target pedestrian SR＿person is increased from 61.15% of the image partition detection to 63.61%.In order to solve the problem of slow processing of model and incomplete detection of normal pedestrian areas caused by image partition detection,a new small target detection method for road pedestrians based on clustering is proposed.Firstly,combined with the distribution characteristics of pedestrian categories,DBSCAN is used to cluster pedestrian category labels to produce a new pedestrian clustering label.Furthermore,a clustering detection model is trained based on Yolo v4,and implement iterative merging and partition filling operations on the clustering results.Then send the adjusted clustering results to the detection model.Finally,when the clustering partition cropping ratio is 0.2,the AP value of the small target pedestrian SR＿person increases from 64.06% when the cropping ratio is 0.2 under the image partition detection to69.51%.In order to enhance the effect of occlusion detection pedestrian,a road occluded pedestrian detection method based on progressive positioning YOLO v4 is proposed.Firstly,the advantages and disadvantages of one-stage network and two-stage network are compared and analyzed.Furthermore,design a new progressive positioning network structure,by adopting the method of stacking detection branches,gradually increasing the detection confidence threshold,so that the anchor can be positioned to the target area more accurately.Finally,when the cropping ratio is 0 under the image partition detection,the AP value of HO＿person and SHO＿person that occluded pedestrian have increased from 33.23% and 17.45% under image partition detection to 37.70% and20.79%,respectively.A driving environment perception platform based on visual image is developed.The different algorithms of this thesis are transplanted to meet the operating requirements of low-configuration computers,and performance tests are performed on different processors.