A Study On Intelligent Vehicle Anti-collision Warning System Based On Computer Vision

With the improvement of the economic level,the number of domestic car ownership has also increased.While cars provide convenience for people to travel,the problems caused by frequent road accidents are attracting more and more attention.According to research literature,warnings to drivers before an accident can effectively reduce the accident rate.Therefore,Advanced driving assistance system represented by the anti-collision warning system is particularly important for reducing the accident rate and improving driving safety.And computer vision takes on the responsibility of perceiving the environment by advantage of its low cost and rich information.This thesis is based on computer vision technology to research and implement vehicle detection and vehicle type recognition,monocular vision.The convolutional neural network is used to realize vehicle detection and vehicle type recognition,and the vehicle distance measurement model based on data regression modeling.The main work of the thesis is as follows:The vehicle type recognition system was improved based on the YOLOV2 target detection model.On the established vehicle type recognition data set,the vehicle types were divided into Car,Bus,and Truck.The Anchor size of this data set is determined by K-means clustering.The network structure of the YOLOv2 target detection framework was improved,and the dense connection network DenseNet-121 was replaced by the original Darknet-19 network as a feature extraction module of the improved model.Tests show that the improved YOLOv2-DenseNet network has increased by 1.36 percentage points compared to the original YOLOv2 mAP.For the positive and negative sample imbalance problem in the One Stage model recognition system,the loss function is optimized by Focal Loss.Reduce the weight of easy-to-classify samples and focus more on difficult-to-learn samples in the defined loss function.In the dataset used in this project,the final YOLOv2-DenseNet-Focal model increased by 1.54 percentage points compared with the original YOLOv2 model mAP,and the final mAP of the model reached 94.89%.The experimental results on the dataset show that while increasing the number of feature extraction network layers,the fusion of low-level features and high-level features is beneficial to improve the detection accuracy of the vehicle type recognition system.The loss function of Focal Loss optimization can solve the vehicle identification model in this subject.The positive and negative sample imbalance problem improves the accuracy of the vehicle identification system.For different vehicle type,based on data regression modeling to achieve distance measurement and design an early warning system.Based on the principle of camera imaging,the vehicle's distance is fitted by acquiring the pixel width of the vehicle under different distance imaging images.The general data regression model only uses the pixel value as the independent variable for ranging.On the basis of the vehicle type recognition system,this paper adds the independent variable vehicle type as a function variable,and obtains the distance regression equation D=F(Type,Pix).Based on the analysis of braking distance,the vehicle distance warning system determines the minimum safety distance and the critical safety distance to realize the multi-level early warning design of the early warning system.The results show that the matching equations for different types of vehicles are more in line with the actual road ranging conditions,thus controlling the ranging error.The multi-level warning strategy is more reasonable and more suitable for urban traffic.