Research On Front Vehicle Detection Methods In Unmanned Vehicle Visual Navigation

In recent years,autonomous land vehicle system(ALV)based on machine vision has become an important content in the current academic research,in which visual navigation technology has attracted much attention.The main content of this paper is based on the vehicle detection research in visual navigation of ALV.Based on the predecessors,further research is carried out.The main contents include:(1)First we design a generic target detection method based on color attribute.This method extracts areas which may contain target in the form of rectangle.We first compare the advantages and disadvantages of the traditional sliding window method and the generic target detection method,and then transform the RGB image into eleven color image by color attribute.Then,we divide the image into different regions.Finally,the color and texture similarity between different regions are used as the evaluation,and the different regions are merged by using the merging strategy.And the merged result is used as a candidate window for the generic target.The experiment shows that the method can describe the candidate regions of the target with fewer windows while maintaining a high target detection rate and has better versatility.(2)The vehicle detection method based on AdaBoost algorithm is studied,then we propose a new method which based on generic target detection and AdaBoost algorithm.We first use the result of generic target detection as the candidate window of vehicle detection,then extract the Haar feature of the vehicle and use AdaBoost algorithm to train the strong classifier,and then generate the cascade classifier as the vehicle classifier to filter the candidate windows.Finally,we get the result of vehicle detection by combining candidate windows using merging strategy.The experimental results show that the proposed method can effectively detect the vehicles in the road,and it is robust and can satisfy the real-time requirement when the resolution is not high.But when the vehicle appears a little block,this method is easy to produce false-positive and missed inspection phenomenon.(3)The vehicle detection method based on HOG-PCA and DPM is studied.The machine learning method based on vehicle feature has poor performance when the vehicle appears a little block and then we study the detection algorithm which based on component.We first extract the HOG feature of the vehicle,then use the PCA algorithm to reduce the HOG feature,and construct the HOG feature pyramid,and then use Latent-SVM to train a single DPM model.Finally,we get the vehicle detection model by combining single Model.The experimental results show that the vehicle detection method based on components is better than the method based on feature when the vehicle is partially blocked.At last,this paper compares the results of AdaBoost algorithm and DPM algorithm,and analyzes the advantages and disadvantages of the two methods.