Research On Small Object Detection And Tracking Algorithm Based On Machine Learning

In recent years,with the development of visible light shooting devices and shooting distance,the shooting environment become increasingly complicated.Small objects with less pixels or features as well as the tremendous similar existing background interferences that result in the detection and tracking problem with small objects tend to be a challenging problem.Regarding the above issues,this dissertation analysed three problems in depth,namely is the detection and tracking of small objects,image pre-processing and small target detection and small target tracking.Through analysing the relevant algorithms,we proposed some algorithms based on its advantages and disadvantages as follows:Firstly,For the small objects image pre-processing,we first studied the traditional morphological-based pre-processing algorithms,and found the limitations of the multistructure open operation reconstruction algorithm,that is the open operation can remove the bright points in the image,and the closed operation can remove the dark points in the image.Based on the closed operation and the multi-structure open operation,this dissertation proposed an image pre-processing algorithm for visible light images which is multi-structured closed operation reconstruction algorithm.Experiments show this algorithm can reduce false targets when the missed detection rate is low.Secondly,For the detection of small objects,it is divided into two parts: regional proposal and classification.In terms of region proposal,due to the reason that the traditional selective search algorithm extracts too many false targets,it is not suitable for small targets.Therefore,Harris algorithm and Selective Search algorithm are combined to make it to be a dedicated small target region proposal algorithm.Experiments show that the algorithm can reduce a large number of false targets and has good robustness to various backgrounds.In terms of classifiers,because the deep random forest algorithm's Multi-Grained Scanning layer cannot extract deep features,it is replaced by a deep multi-grained scan layer,and experiments shows a better accuracy compared to the former algorithm.Finally,For small objects tracking,the traditional objects tracking network siamese network was analysed.Since the siamese network uses improved AlexNet as a feature extraction structure,which means this network contains dimension reduction operations that will result in the upper layer feature of the small-target feature may completely disappears.Therefore,multiple small convolution kernels are used instead of the large convolution kernel to achieve sufficient receptive field,the dimensionality reduction operation is removed as well to retain more small-target features.Experiments show that the algorithm can effectively track small objects and have a certain robustness to the complex background.