Moving Object Detection Based On Imbalanced Learning

Recently,electronic devices with smart cameras have shown up enormously.These devices carry fast and efficient computer vision applications that include moving object detection strategies.Moving object detection is a binary classification,which aims to classify a video into two classes: foreground(moving object)and background.It plays an important role in public security,traffic monitoring and so on,and it is also the basic step of follow-up tracking and recognition.Model initialization,illumination changes,dynamic background and other factors affect the accuracy of segmentation,but the sample set used to classify is imbalanced actually,which is ignored.In the case of binary classification,class imbalance means that the number of negative samples is greater than the positive ones.Correspondingly,in moving object detection,the background samples are negative(majority class)and its number is larger than the foreground samples which are positive(minority class).By improving the imbalance level of the data set,the accuracy of detection can be effectively enhanced,therefore moving object detection based on the imbalanced learning is worth of study.In this dissertation,the problem of moving object detection based on imbalance learning is considered from the data level and the algorithm level.The main contents are as follows:1.The data-level solution is resampling.On the one hand,minority groups are over-sampled to generate new samples,and the instants of over-sampling is designed according to certain rules to guarantee that the newly synthesized samples are evenly distributed within the sampling interval,thus preventing over-fitting.On the other hand,the negative samples are selectively down-sampled,and the frame differencing is executed between background frame and one synthesized frame.Hence the number of background samples located in the overlapped area between background and foreground are reduced.Through the above two steps,the imbalance degree of data is decreased,getting a balanced data set for classification.Finally,the feasibility of the algorithm is verified via experiments and the accuracy of the detection is improved.2.The algorithm-level solution is cost sensitive,that is to say,applying varying misclassification penalties to the positive and negative samples.Different to the widely utilized constant costs,two pixel-wise cost functions are constructed.In the MAP-MRFframework,diverse misclassification penalties are assigned to foreground and background respectively to strengthen the accuracy of the models.Data sets obtained from the previous resampling are employed to the classifier based on cost functions.Experiments demonstrate that the combination of more imbalanced compensation strategies can further improve the integrity of foreground.