Research On Optical Object Detection Technology In Complex Background

As an important problem in artificial intelligence and computer vision,object detection can be widely applied in various applications including video surveillance,driver assistant systems and robot navigation.In this paper,we focus on the research of classic moving object detection technology and object detection technology based on deep learning.Particularly,we focus on the ViBe and R-CNN methods and achive some research results.The main work of this paper is concluded as follows.We analyze the optical object detection technology.Firstly,the classic moving object detection algorithm is introduced in details,and the frame difference method,the background subtraction algorithm and the optical flow method are respectively analyzed in theory together with their merits and demerits.Then,we introduce the Convolutional Neural Network(CNN).Deep convolutional nets have brought about breakthroughs in images,video,speech and audio processing.We describe the deep convolutional nets in details from the aspects of network structure,basic idea,theoretical derivation,merits and demerits,which explain why CNN has a good performance in object detection.We study a universal background subtraction algorithm for video sequences(i.e.,ViBe).Firstly,we introduce the ViBe algorithm,and then analyze the disadvantages of ViBe algorithm in principle.Aiming at the problems of ghost,background turbulence in high frequency,camera jitter and error of background update caused by spatial propagation technique in the classic Vibe algorithm,an improved ViBe algorithm is proposed.The new method decides whether the ghost target exists in the background model or not,combining with visual saliency,and adaptively changes the time subsampling factor through the level of ghost for each pixel in the background model,which can improve the rate of ghost elimination.Self-adaptive threshold is adopted in the process of model matching by establishing a blinking degree matrix to check the high-frequency disturbance level of background,so that background models is better suitable for the dynamic background.Small object discardisng and hole filling strategies are added to the new method.It determines whether a foreground pixel is a noise point caused by camera jitter or an error of background update by counting pixel numbers in 24-connected neighboring region of foreground pixels.Since that,it improves the robustness of the algorithm.Experiments demonstrate that the improved algorithm is a good way to make up for the deficiency of the original ViBe algorithm.The accuracy and recognition rate are greatly improved.We study a region-based convolutional network((R-CNN))for accurate object detection and segmentation working on a single image.Firstly,we introduce R-CNN in details from three aspects: algorithm framework,algorithm analysis and algorithm implementation and design details.Then,based on the shortcomings of current R-CNN,which has poor performance in small object detection and has a training of multi-stage pipeline,an improved R-CNN based on end-to-end training combined with multi-scale representations is proposed.And we detail the principles of improvement.Finally,we test the improved R-CNN in test images and VOC 2010 datasets,the results show that the improved algorithm achieves good performance for small object detection because of the integration of the multi-scale features.Moreover,a further improvement can be found in mAP for multi-task joint training and two bounding box regression.