Features Learning And Real-time Detection Of Natural Objects Based On Machine Vision

Features learning and real-time detection of natural objects based on machinevision is crucial for many applications both in theory and in practice. Focusedon this two area, the thesis researched some challenge problems and contributedto it with the achievements of generalization. To the features learning of naturalobjects, the main achievements are listed below:1. Experiments were done on adaptive features learning of color-reduction,which in fact is a processing of clustering. This method can be used to extracttextural pattern and is propitious to promote the output quality. Adaptive E-Mclustering of 1D or 2D Gaussian-mixture model, principal curve learning withself-cross data clouds, are of the function with learning guided, however, theoutput speed is quite low (all below 1500ms).2. A real-time visual surveillance method, especially for intruding alarmand tracking, was presented, which is based on macroblock features quantificationof competitive learning. Critical system parameters, such as threshold andprocessing regions, are adaptive to varying frame rates and computationalconstraints. This method is effective and efficient in eliminating the regionnoise caused by fluorescent lamp, local moving shade and dithering in dynamicimage sequences. Several experimental results, exhibited the promisingperformance in real-time visual surveillance in a changing environment.3. A novel method of defects detection of surface was presented, which isbased on non-supervised roll-style multi-features learning and enhancement inspatial-temporal domain. So that the classfication evidences extracted are moresolid and reliable.4. A novel extended FloodFill algorithm was presented, which combined witha ring-style learning pattern of RGB color features to archive the ideal kernelregions in natural field scenario or interested textural objects, where theforeground and background is of similar colors. The best coefficient and operatorof color components is determined by a supervised learning mechanism accordingto related special color space, with a computation cost less than 10s to an imageof 1280×1024 resolution. By this method, some crucial off-roads were extracted.5. To the features learning of natural non-rigid and semi-transparentobjects, this thesis focused on the research of video-based early smoke detection.A novel two-level controlled background-learning model was proposed, where thebackground is updated via iteration between temporary and long-term background.The achieved results of 4 natural dynamic video sequences (about 1443 frames forhighway, 1370 frames for smoke) shows that the proposed method of backgroundlearning is promising and closed to practical application.To the fast detection of natural objects, the main achievements are listedbelow:1. Focused on automated recognition and counting of colony or seeds, an counting algorithm of ellipse model was presented, with a correct check out rateup to above 96.5% to 40 images, on which complex objects distribution wereformed. The count out strategy can reduce the disturbing influence among objects.2. Evidence enhancement in spatial-temporal domain were used in surfacedefect detection, which is carried out with non-supervised techniques. The mainidea is to check out the unknown defects regions based on the multi-features ofthe known defects-free textile. Simulating results of on-line inspection showthat the efficient detection speed reaches 55 frames per second to a concolorousimage sequence (1024×393 pixels) of the unknown textile, with a correct checkout rate of surface defects up to above 95% among 7 sequences.3. Active vision on application of agricultural field, especially with theboundary tracking of cut-uncut crop surfaces with similar colors, is quite achallenge. Two novel methods, i.e., the multi-evidence fuzzy enhancement frompixel-rows (MEFE) and the multi-rows Best Fit Step (MR-BFS) were proposed forfast segmentation of video sequence in order to navigate agricultural robot. Thetwo approaches are nearly non-supervised and their output of guideline is ableto be adaptive to a changing environment in somewhat. The achieved results of5 dynamic video sequences (about 1200 frames) shows that correct segmentationwas achieved an average error below 5% for normal sequence, where segmenting byMEFE is more faster (correct segmentation is done within 70ms for color sequenceof 320×240 resolutions) and that by MR-BFS is a final trade-off.4. A real-time method of projective transformation and an simpleauto-calibration method for camera's main pose (with a time cost less than 0.5s)were presented for online boundary tracking of cut-uncut lawn. Dynamic weightedaverage among history guiding angles and segmenting positions of video sequencecan keep the tracking decision in a reasonable continuity, with may reduce theaction vibratility.5. A novel real-time technique for segmenting image into regions of interest(ROI) of field scenario in active vision was presented, such as tracking croprows or rural road, extracting crops or weeds, colony or seeds, which is basedon linear combination of RGB components and their bit-masked color reduction.This method can be used to.detect early smoke and to deal with the draft extractionof textile image too. The achieved results of 6 dynamic video sequences (about2300 frames) shows thatthe processing speed is influenced little by texturalstructure of regions, with full segmentation of multi-objects being done within30ms for color image sequence of 320×240 resolutions, and is robust against thevariation of color noisy distribution and illumination in some limits.6. Combining with recent smoke recognition technologies, the regions ofinterest like smoke event were extracted through the multi-evidence of colorhomogeneous features, color tolerance learning, wavelet features, smoke movingfrequency analysis and smoke's semi-transparent feature. Smoke sizes wereprecisely figured out by isolated region control. Connected components labelingwas also implemented for alarming treatment. The achieved results of 3 natural dynamic video sequences (about 1370 frames for smoke detection) shows that theproposed method of background learning and early smoke detection is promisingand closed to practical application.