Research And Implementation Of Vision-based Detection Methods For Vehicles In Front Of Unmanned Vehicles

With the development of artificial intelligence technology,Internet technology and automobile industry,the degree of automobile intelligence is gradually increasing in recent years.Unmanned technology has become a hot research topic at home and abroad.Vision-based vehicle detection has received widespread attention as one of the most important technologies.The differences of vehicle appearance,the impact of light and the complexity of the driving environment are difficulties of vehicle detection.This article mainly studies vision-based vehicle detection methods and applies them to unmanned vehicle platform.An improved geometric constraint algorithm is proposed to obtain candidate regions for vehicle detection in this paper.Firstly,the world coordinate system and the image coordinate system are established,and the transformation matrix is calculated.Then,different sizes of sliding windows are designed for detecting targets of different sizes.Then according to the geometric constraints between the size of sliding window and the vehicle in the original image,candidate areas that may contain the vehicle are obtained in the original image.Based on the y-direction constraint in the image coordinate system,the method of this paper also constrains the vehicle's candidate region in the x-direction according to the left and right lane lines.Finally,the candidate areas are mapped to different resolution of the scaled image.This method avoids the repeated calculation of the matrix compared with the original algorithm by mapping the candidate regions to different resolution images.And the method of this paper constrains the candidate area according to the left and right lane lines,further improving the sliding window efficiency.Target detection based on sliding window is a simple and effective detection framework.Based on the improved geometric constraint method,two vehicle detection methods are proposed:a multi-feature fusion method based on confidence and a CNN-based verification method.(1)Vehicle detection method based on multi-feature confidence fusion.First,Haar-like features,HOG features,and LBP features of the image are extracted.Then,the three features are input to Adaboost classifiers respectively to obtain three classification results and the corresponding confidence values.Then the three confidence values are normalized according to their distribution of training samples.Three confidence values are fused to obtain a new one using the weighted fusion method.Finally,the fused confidence value is used for vehicle detection.Experimental results show that this method can improve the accuracy of vehicle detection effectively.(2)Vehicle detection method based on CNN verification.CNN can self-study the intrinsic characteristics of the image with enough samples.Cnn-based target detection method has a higher detection rate and slower detection speed.Thus,Our method first calculates the LBP feature of the image.The LBP features are input to the AdaBoost classifier.The classifier quickly excludes most areas that do not contain vehicles.The classifier then outputs candidate rectangles that may contain vehicles.Finally,a vehicle classifier based on CNN model is designed to validate candidate rectangles.The experimental results show that the proposed method is robust to light,attitude of vehicle and partially obscured vehicles.It can meet the real-time requirements and achieve higher detection rate and lower false detection rate.Finally,this paper introduces the system components of unmanned vehicle platform,focuses on the software and hardware environment of vehicle detection,and applies the two vehicle detection methods proposed in this paper to the unmanned vehicle platform.The experimental results show that the detection accuracy of the vehicle detection method based on CNN verification is 97.72%,the recall rate is 95.39%and the average processing speed is 257 Ms/frame.This can meet the real-time system requirements.