Research And Application Of Target Detection Based On Decision Tree

Target detection is one of the important contents of computer vision research.The research results of target detection have been widely used in the fields of intelligent video surveillance,robot navigation and image retrieval.The target detection in specific situations and different environmental conditions not only needs to compromise the complexity and reliability of the algorithm,and sometimes consider whether there is real-time.With the popularity of on-board surveillance cameras and the development of computer vision technology,many researchers have begun to try to use the target detection technology to make statistics to bus passengers.Quadratic tree algorithm is often used in the target detection,with the advantages of simple generation,high classification accuracy and good robustness to the noise in the data.This paper studies the target detection technology based on quadratic tree algorithm and applies it in the bus environment.The work done in this paper mainly includes the following aspects:First,the paper propose cost-sensitive quadratic tree algorithm.The face or head relative to the entire background is a small probability event,for such a small probability event,it can be considered the price of missing a positive sample window is more expensive than false detecting a positive sample window.So cost-sensitive learning has been introduced.The algorithm extracts the Normalized Pixel Difference(NPD)features to train the model.Then,the extracted features are sent to deep quadratic tree,and the deep quadratic tree is used as the basic classifier to train cost-sensitive Gentle Adaboost target detection algorithm..In this way,a number of deep quadratic tree are established and then cascaded with Soft-Cascade to obtain the final cost-sensitive decision tree target detection algorithm model.The experimental results show that compared with the existing deep quadratic tree algorithm,the proposed algorithm improves the face detection rate and detection speed on the FDDB data set and bus video.In the bus environment,head detection,crowding scene detection rate was 92.8%,non-crowded scene detection rate of up to 98.0%.Second,the paper combined with the maximum correlation minimum redundancy algorithm to optimize the deep quadratic tree target detection algorithm model.Due to the large number of NPD features,the features are selected by the maximum correlation minimum redundancy algorithm before the deep quadratic tree begins to learn.Maximize the correlation between features and categories,minimize the correlation between features,select the best subset of features,and use the optimized subset of features to train the deep quadratic tree.Experiments show that,adding this method can optimize the model,the detection rate is almost unchanged,the detection time increased by 30%.