Small Obstacle Detection In Automatic Driving Scene Based On Machine Vision

In recent years,driverless technology has developed rapidly.At present,driverless vehicles have been tested on the road at home and abroad.In the future,driverless vehicles are likely to appear on urban and rural roads at home and abroad on a large scale.The current research and development tests mainly focus on ensuring that the vehicle reaches the destination safely and smoothly.On this basis,the comfort of the vehicle will naturally become a new demand in the future.Small obstacles in the road,such as bricks and stones,speed bumps,potholes,etc.,may not affect the safety of unmanned driving,but will cause vehicle bump vibration,which will inevitably affect the stability and ride comfort of the car during driving.Taking bricks,speed bumps and road potholes as examples,this thesis studies the detection of small obstacles that affect the stability of vehicles on the road,and hopes to detect and avoid them in real time in the future.Firstly,based on the advantages and disadvantages of traditional edge detection algorithm,this thesis proposes an improved Canny edge detection algorithm.The proposed improved median filtering algorithm is used to replace the Gaussian filtering in the original Canny algorithm for image smoothing.Since the double threshold in the lag threshold processing of traditional Canny edge detection is artificially set by experience and globally fixed,the algorithm lacks adaptability.In this thesis,the sliding window combined with Otsu algorithm is used to automatically set the double threshold for Canny edge detection to improve the adaptability of Canny algorithm.The improved Canny algorithm is used to extract sub-images that may be target obstacles in the original road image.Secondly,this thesis designs a DT-SVM(Decision Tree SVM)classifier to classify and identify sub-images that may be target obstacles.The HOG(Histogram of Oriented Gradient)feature and LBP(Local Binary Patterns)feature are improved and fused to obtain the HOG-LBP joint feature extraction algorithm to improve the accuracy of the classifier.Aiming at the problem that the traditional SVM(Support Vector Machines)algorithm is only suitable for binary classification,a SVM based on decision tree is designed for multi-classification tasks.Firstly,the data are classified according to the principle of class priority separation with the maximum minimum distance between all other classes.When the distance between multiple classes is equal,the data are classified according to the principle of class priority separation with high probability of occurrence.Ensure the maximum accuracy,while reducing the number of classification judgments as much as possible,shorten the classification time of the classifier.Finally,based on the perspective transformation,the obstacles that affect the driving are determined to determine whether the target obstacles identified by the DT-SVM classifier will affect the driving of the vehicle.First,camera calibration is performed to eliminate distortion,and the original road image is converted into a bird ’s-eye view form through perspective transformation.Through the relationship between the driving path of the vehicle in the aerial view and the position of the target obstacle identified by the classifier,it is judged whether the target obstacle affects the driving of the vehicle and completes the final detection task.This thesis mainly studies the detection of small obstacles that affect the stability of unmanned vehicles on the road,detects small obstacles in the road and determines whether they affect the driving of the vehicle,assists the unmanned vehicle to avoid obstacles or slow down,thereby improving the comfort of the unmanned vehicle.