Research On Pedestrian Detection Technology Of Mine Locomotive

Mine locomotives are the main means of transport in underground shafts.As the demand for mineral resources continues to rise,the transportation tasks of mine locomotives are becoming more and more heavy.The locomotives are prone to collide with pedestrians in complicated underground environments.Therefore,It is of great significance to ensure the safety of underground miners.At present,at home and abroad in this regard to do a lot of related research,although achieved some results,but in practical applications,still face many problems.This paper analyzes the present situation of pedestrian detection in underground,and uses the image processing technology,combined with the actual project requirements to study the track detection,pedestrian feature extraction,unbalanced sample data and pedestrian classification in the pedestrian area.Specific as follows:1.A method of Hough transform orbit detection based on polar angle constraint is proposed for track misdiagnosis and false detection in traditional orbit detection.This method selects a certain number of high-definition template orbit images,obtains the appropriate polar angle and polar range by traditional Hough transform,and limits the voting area of the traditional Hough transform parameter space,and finds the local peak in the parameter space after the constraint.Using this method not only can improve the accuracy of track detection,but also reduce the amount of calculation.2.A low dimension pedestrian feature extraction method is designed.This method combines the cell elements of the original HOG feature with the block units to reduce the image redundancy characteristics.The number of sliding blocks of the HOG feature detection window is reduced,and the look-and-watch method is used in the statistical gradient histogram.The experimental results show that the method can reduce the HOG feature dimension in an efficient way.3.A sample feature space segmented oversampling method is proposed to balance the data distribution of the training samples.Due to the particularity of the underground environment,it is often difficult to collect a large number of pedestrian training images,which will make the training sample data distribution uneven problem,affecting pedestrian classification accuracy.In order to solve this problem,this paper proposes a segmented oversampling method based on the analysis of the traditional oversampling algorithm.The feature space of the minority samples is segmented.For the surroundings of the decision function and other minority samples with the classification value area,Respectively,the use of different oversampling method to generate a new number of samples,through this method can effectively alleviate the training sample exists in the data distribution imbalance problem.4.In order to improve the accuracy of pedestrian classification,this paper improves the pedestrian classification method based on AdaBoost and SVM combination.This method integrates the SVM algorithm into the AdaBoost algorithm framework by using the powerful combination ability and SVM algorithm of AdaBoost algorithm in the characteristics of small sample data learning ability.By classifying the training set in order,the SVM weak classifier model is obtained,and the optimal combination is selected from these weak classifiers to form the final strong classifier.Compared with the traditional AdaBoost classifier and SVM classifier,The classification accuracy of the method has been improved.