The Research Of Moving Objects Detection Based On Convolutional Neural Network

As one of the most important and basic tasks in video analytics applications,motion objects detection has always been regarded as a practical and challenging topic in the field of computer vision.Although a large number of works on this issue have been published in recent years,due to the complexity of natural scenes and random noises,there is still a lack of a universal and efficient solution.Traditional moving objects detection algorithms focus on establishing a statistical background model to segment moving objects by finding stable features.However,artificially designed features often fail to adapt to the challenges of complex natural scenes,such as changes in lighting,target occlusion,dynamic background,and so on.In contrast,deep learning methods,which start to flourish in the field of computer vision,do not require researchers to manually design features and can be robust and adaptable in many complicated scenarios.Thus,the Convolutional Neural Networks will be used as a tool for learning features in this paper.In order to overcome the shortcomings of traditional methods in complex scenes,this paper focus on the pixel variations and the temporal distribution of pixel values,and establish two motion detection methods respectively.The main works are as follows:(1)In the method which is called Deep Pixel Variation Learning(DPVL),a feature called Pixel Patch is proposed to encode the pixel variation.And an end-to-end full convolutional neural network is designed to learn the patterns of pixel variation at each pixel position,and finally the value of current pixel is transformed into a subspace which guarantee the linear separability for classification.In addition,some neighbors of the current pixel are randomly sampled to obtain the additional Pixel Patch,so that our network can obtain the constraint of the spatial context,thereby improving the classification performance of the DPVL method.(2)In the method which is called Deep Pixel Temporal Distribution Learning(DPTDL),a Comparison of Random Pixels feature(CRP)is proposed for encoding the temporal distribution of pixel observations.And in this paper,a convolutional neural network is designed to learn the temporal distribution of pixel observations.By comparing the difference between the current pixel and its historical observations,it is determined whether it belongs to the foreground or background.Compared with traditional motion detection methods,the proposed detection methods are not limited to the simple pixel models and artificially designed features,thus they are more capable to adapt to the complicated scenarios in the real world.Compared to other methods of motion detection based on deep learning,their background models still rely on traditional methods,which limits the effectiveness of motion detection.The two detection models proposed in this paper directly model the pixel observations,avoiding the limitations imposed by the above methods.Comprehensive experiments on several standard benchmarks demonstrate the superiority of the proposed approach compared to state-of-the art deep learning and traditional methods.