Dynamic Pedestrian Or Vehicle Recognition And Range Estimation Based On Deep Learning

The two important indicators of object detection and range estimation in front of the vehicle are real-time performance and accuracy.Therefore,high real-time performance and high accuracy are the prerequisites for automatic driving.Vehicle and pedestrian detection in the driving environment is an important basic task of automatic driving.Due to the complexity and dynamics of realistic driving environment,the vehicle and pedestrian detection methods based on statistical learning can not meet the needs of automatic driving in the new period.Therefore,the research of high-performance detection algorithm becomes a main task in the future.At present,the object detection algorithm based on deep convolutional neural network has been proved to be a learning method that can effectively describe the semantic mapping of regions,which provides a new approach for object detection research.This thesis explores the application of deep learning algorithms in image object detection in recent years and establishes an improved YOLOv3 object detection model.At the same time,this thesis also combines the traditional geometric range estimation model and mathematical regression model to build a neural network model for ranging based on object detection framework.This thesis mainly studies the algorithm of object detection and ranging estimation in the driving environment based on deep learning.The specific research contents and conclusions are as follows:(1)For the task of object detection,this thesis takes advantage of the algorithm YOLOv3 with high accuracy and real-time performance.Then,considering the particularity of driving environment,the following three improvement strategies are established: First,the K-means algorithm is used to calculate the most optimal anchor box parameters;second,the model is trained in two stages based on the weights of the Darknet-53 network;third,the model is trained after balancing the number of categories based on the combination of oversampling and undersampling.To some extent,the performance of the model is affected by the size of the training data set.Therefore,this thesis collects a large number of image data sets from KITTI,and processes the corresponding labels of each image.Finally,a reasonable training data set is established.The object detection model is generated by training the improved YOLOv3 algorithm,which provides a basis for the design of the object ranging model.(2)For the problem of range estimation in front of vehicles,this thesis analyzes thecharacteristics of object range distribution in KITTI data set,and build ranging models for different categories,and then integrates them together.Due to the shortcomings of traditional ranging algorithm,such as camera calibration,pitch angle measurement and so on,two ranging schemes are established in this thesis: one is to build a monocular ranging model by using the data regression idea and the pinhole imaging principle,the other is vehicle ranging algorithm based on deep neural network.Finally,combining the advantages and disadvantages of the two schemes,the range estimation of object in front of the vehicle is completed.(3)The feasibility of the designed system is verified by building an experimental environment.The experimental results show that the object detection system established in this thesis can accurately detect the vehicles and pedestrians in the driving video in a real-time way,and can calculate the distance between the current vehicle and the surrounding objects through the ranging model.