Research On Occupant Detection Methods For Smart Airbag Application

Airbags into automobiles have significantly improved the safety of the occupants.Unfortunately, airbags can also cause fatal injuries when vehicle collision. Automotiveairbag, which can detect occupant classification, position, collision types etc. to control theairbag stress and velocity, will be used to prevent occupant re-harmed by airbagdeployment. Accordingly, the occupant type and position one of the important part ofautomobile intelligent constraint system are key factors to decide the stress and velocity ofairbag to deploy.To detect occupant type and position, the methods of occupant detection for smartairbag are proposed and hardware development platform based on vehicular multi-sensorfusion is developed and the methods on occupant classification fusion decision andtracking is putted forward in the Thesis. The mainly research work is presented as follows:1. The occupant detection system and Hardware platform is developed for smartairbag.According to the characteristics of the occupant detection system, the Hardwareplatform is developed by TMS320DM642digital signal processor (DSP), video and seatpressure sensor etc. and the occupant multi-sensor fusion detection software framework ispresented by occupant classification and occupant tracking and motion estimation. Forlacking occupant test video sequence for public testing, lots of videos under differentcircumstance on occupant are captured as testing data source for occupant detection2. The occupant image quality evaluation method on structure similarity of gradientdirection is put forward to obtain high quality occupant image for detection.It is difficult to detect occupant image when the occupant image is blur or unevenbrightness. To overcome the difficulty, the quality evaluation method on the structuresimilarity of gradient direction which is robust to sunlight, illumination intensity and samehuman visual system is proposed in the thesis. The structure similarity of brightness,contrast and direction of the gradient is computed by local regions of the reference imageand current image to evaluate occupant image good or not in the method.3. Contourlet transform is used to reduced noise and local equalization method is put forward to enhance the occupant imageImage noise and illumination is the important factors to restrict the occupant imagequality. To overcome this aspect effects, the method about reducing noise by Contourlettransform to occupant image and image equalization enhancement about illuminationestimation is put forward. As the Contourlet transform can realize multi-scale arbitrarydirection on any scale decomposition, better keep the outline of image and texture directioninformation and make up for the inadequacy of the wavelet transform, in this paper. Thismethod got the Contourlet coefficient of the occupant decomposed images, which useContourlet transform, control the Contourlet coefficient of the high frequency of theimage region, which use coefficient threshold and threshold function inhibition, andcombine the reconstruction of occupant images with Contourlet inverse transform to realizeoccupant image noise reduction processing. And then, to reduce the influence of imageintensity by suppressing the DC component of the image of Fourier transform and map thelight intensity to the [0,255] the brightness of the space by linear stretching to the lightintensity. In the last, the processing of the occupant image enhancement is realized byenhancing the local image by using the method of the partially overlapped sub-block localhistogram equalization, improving the image brightness distribution, balancing theoccupant image, and increasing the contrast of image details4. Occupant region estimation algorithm using the gradient direction histogram is putforward to estimate the occupant regionOccupant region estimation can reduce the search range in occupant detection, and itcan reduce the interference by edge of no-occupant objects. For occupant image isinfluenced by sunshine dynamic projection, shadow, the landscape outside the window ofcar, other matters inside, etc, when the vehicle is moving, It is difficult to get the occupantregion by using the method of background difference for dynamic image detection andframe difference. Based on this, the feature expression of local characteristics of the localorientation histograms is set up to reduce sensitive degree to illumination, noise andcontrast in this paper by using the mutual spatial structure characteristic between occupantimage sequence and the background of the occupant. To set up local image mappingrelation table by using the cosine theorem to metric vector angle and express the relativedegree of vectors, get occupant area and realize occupant area estimation by using the splitand merge process of local area mapping relation table and getting rid of local block ofisolated nodes.5. The occupant multi-sensor fusion classification method is put forward to classifythe occupant type. It is difficult to detect occupant type and position accurately by single sensor for the occupant type, sunshine, and so on. In order to overcome the problem, in thispaper the multi-sensor fusion decision recognition method using the video sensor andpressure sensor is presented, the multi-sensor fusion strategy was designed, and theequipment layer and decision layer fusion strategy were determined. At same time, windowclassifier framework based on occupant HOG and SVM, built multi-scale occupantwindow classification fusion location method to realize the occupant video detectionmethod. According to the pressure information of pressure sensor, it built occupantclassification decision method for empty, and according to the decision layer fusionstrategy of the video and pressure sensor, the method of multi-sensor fusion decisionclassification was proposed, which can realize the classification of the empty, adults andchildren.6. Motion estimation method of occupant tracking based on semi-supervised learningand IMM was proposed.In order to meet the rapid reaction on automobile safety airbag and locting occupantposition, according to the characteristic of feature change and multimodal movement whenoccupant is tracked, occupant tracking with feature update and IMM based onsemi-supervised online learning and motion estimation method was set up in this paper.The method used the thought of detection before occupant tracking. Firstly, we get theoccupant feature selecting strong and weak classifiers for occupant off-line using adaboostframe, according to the occupant area,which is decided by current static occupantclassification window and positive and negative samples, which is decided by thebackground, realize online update selecting for occupant feature and get the prior classifierof occupant classification. Secondly, according to the prior knowledge, the classifierdetected occupant, using Mean shift kernel function estimates the similarity of spatialdistribution of occupant classification metric. And then the metric as occupant tracking arearegions of maximum was got. Thirdly, according to the strong and weak classifiers foroff-line occupant and the weight distribution, realize the semi-supervised online learningalgorithm with the online update occupant feature selection using the un-labeled ofoccupant area tracked at present. Finally，realize the method of multi-mode occupantmotion estimation by using IMM frame and multi-mode Kalman filter, via mutual input,mode filter, update probability and combination estimation. This method meets theoccupant tracking process appearance characteristics changed characteristics of theoccupant online update and occupant movement of multi-mode tracking and motionestimation, and restricts the drift phenomenon in online Boosting.Videos under different circumstance such as occupant type, illumination and occupant state and so on were tested by proposed methods, and the results showed the proposedmethod can meet the need of occupant classification and predict the position of occupantfor smart airbags application, meanwhile the proposed methods can be as a reference forfuture research in depth on smart airbag.