CN105574511B - Adaptive object classification device with parallel framework and method thereof - Google Patents

Abstract

Translated fromChinese

本发明公开了一种具平行架构的适应性物体分类装置及其方法，其中具平行架构的适应性物体分类方法包含储存多笔场景参数及多笔分类器参数的步骤，先撷取至少一影像数据后，框选出影像数据中的多个障碍物影像，可根据每一障碍物影像范围选定适当的框选范围。再利用多个影像处理单元以一对一的平行处理方式分别对应计算出多个障碍物影像的多笔障碍物特征数据，并根据多个障碍物特征数据选取对应的场景参数及对应的多个分类器参数并进行运算，以取得多笔分类数据，若判定分类数据为所预设的障碍物种类，则输出所框选障碍物影像的位置，如此可达到实时侦测障碍物的目的。

The present invention discloses an adaptive object classification device with a parallel architecture and a method thereof, wherein the adaptive object classification method with a parallel architecture includes the steps of storing multiple scene parameters and multiple classifier parameters, firstly capturing at least one image data, then selecting multiple obstacle images in the image data, and selecting an appropriate selection range according to the range of each obstacle image. Then, using multiple image processing units to respectively calculate multiple obstacle feature data of multiple obstacle images in a one-to-one parallel processing manner, and selecting corresponding scene parameters and corresponding multiple classifier parameters according to the multiple obstacle feature data and performing calculations to obtain multiple classification data, and if the classification data is determined to be a preset obstacle type, then the position of the selected obstacle image is output, so as to achieve the purpose of real-time obstacle detection.

Description

Have the adaptability object sorter and its method of parallel framework

Technical field

The present invention relates to a kind of adaptability object sorter and its methods, particularly relate to a kind of adaptation for having parallel frameworkProperty object classification devices and methods therefor.

Background technique

Since traffic safety is increasingly taken seriously in recent years, as the cost of image documentation equipment is greatly reduced and image identification skillArt is increasingly mature, and image recognition system is applied more and more extensive on vehicle safety, and in the security system, is distinguished using imageKnowledge is to reduce an important method of whole system cost.It is shown according to statistical data, as long as driver is collidingEarly warning was obtained before 0.5 second, it can avoids at least 60% car accident that knocks into the back, 30% head-on car accident and 50% roadFace related accidents, and if have 1 second pre-warning time, can avoid 90% accident.But in image recognition system, calculation amountHuge is a main difficulty, and real-time operation is required more to need to be taken seriously in harsh Vehicle security system.

For anti-collision system, the pedestrian's detecting system used is usually all very expensive equipment, for example uses infrared rayDetecting, laser radar detecting etc., have multiple pedestrians, vehicle and cat and dog due to the complexity of road scene, such as in Same Scene,Therefore pedestrian and other background informations are separated with greater need for more strong characteristic parameter.In addition, pedestrian's detecting system is being detectdDuring survey, it is subjected to the interference of the various environmental change factors of floor and the accuracy rate of detecting result is declined, such asCause pedestrian part excessive lightness or darkness in the environment of uneven illumination, or in the case where the drive body of pedestrian is partially obscured,Usually whether can not accurately it judge in scene with the presence of pedestrian.The method of another pedestrian's detecting is to utilize background acquisition methodForeground information is obtained using as image treatment method further, however obtains being more broken using the method pick-up imageProspect, cause rear end recognize difficulty, and because need for a long time capture, to increase the burden of system.Therefore, how to mentionIt rises the detecting efficiency of barrier and meets detecting real-time demand and be a problem to be solved.

In view of this, inventor satisfies in view of the above shortcomings of the prior art, a kind of adaptation for having parallel framework is proposedProperty object classification devices and methods therefor, it is above-mentioned effectively to overcome the problems, such as.

Summary of the invention

The main purpose of the present invention is to provide a kind of adaptability object sorter for having parallel framework and its methods, put downRow handles multiple barrier images, so as to accelerate image processing efficiency, and then simplifies the complexity and reduction of image processingIt is time-consuming.

Another object of the present invention is to provide a kind of adaptability object sorter for having parallel framework and its method,Corresponding scene can be judged according to the image captured, it can be according to various different scenes come elastic adjustment classifier parameters, energyThe higher barrier image of accuracy and its type are calculated, enough to solve the problems, such as existing detecting system erroneous judgement.

Another object of the present invention is to provide a kind of adaptability object sorter for having parallel framework and its method,It can select barrier image capturing range appropriate according to the remote depth of the image captured and the size of barrier image, frame and carry out shadow againAs processing, it is able to ascend image processing efficiency.

A further object of the present invention is to provide a kind of adaptability object sorter for having parallel framework and its method,It is embeddable in the anti-collision system of vehicle, to achieve the purpose that detecting real-time.

In order to achieve the above object, the present invention provides a kind of adaptability object classification method for having parallel framework, the partyThe step of method includes more scenario parameters of storage and more classifier parameters, adaptability object classification method includes the following steps:Capture an at least image data；Frame selects multiple barrier images in image data, and according to each barrier image capturing rangeSelect selection range appropriate.Multiple image process units are recycled to respectively correspond more barriers for calculating multiple barrier imagesHinder object characteristic, and chooses corresponding scenario parameters and corresponding multiple classifier parameters according to multiple barrier characteristicsAnd operation is carried out, to obtain more classification data.Wherein, multiple classifier parameters include more directions of different obstacle speciesGradient Features parameter, each classification data are multiplied by corresponding classifier parameters by each barrier characteristic, and add up andIt obtains.Whether match stop data are greater than a floating threshold, the corresponding obstacle species of floating threshold, if so, determining classificationThe corresponding obstacle species of data, and export institute's frame and select the position of barrier image, appoint if it is not, then determining that classification data is not correspondingWhat obstacle species.Since the present invention not only can effectively shorten operation time with the multiple barrier images of parallel processing, fromAnd achieve the purpose that detecting real-time, and corresponding scene can be judged according to image data, it can not only be suitable according to scenario parametersAnswering property adjusts classifier parameters, and barrier characteristic of arranging in pairs or groups carries out operation to obtain more reliable classification data, therefore canWhole detecting accuracy, great market competition advantage is substantially improved.

Wherein, further comprising the steps of before frame selects multiple barrier image steps: from image data select one have it is emergingInteresting region, then select from interesting region center multiple barrier images in image data.And in view of in image dataMultiple barrier images have remote depth and image size issue, therefore after frame selects multiple barrier image steps, it further include that adjustment is everyThe step of selection range size of one barrier image.

Wherein, direction gradient algorithm is further comprising the steps of: first calculating the marginal value of each barrier image, direction gradientEach barrier image is divided into multiple block of cells by algorithm (HOG), then calculates the direction of each pixel in each block of cellsAnd edge parameter value and add up, to obtain nine feature vectors, and nine feature vectors in each block of cells are counted, is hinderedHinder the barrier characteristic of object image.

The present invention also provides a kind of adaptability object sorter for having parallel framework, which stores more scenesParameter and more classifier parameters, adaptability object sorter include an at least video capture device and an image processor.ShadowAs an acquisition device acquisition at least image data, and frame selects multiple barrier images in image data, and according to each obstacleObject image capturing range selectes selection range appropriate.Image processor includes multiple image process units, multiple image process unitsMore barrier characteristics for calculating multiple barrier images are respectively corresponded, according to the selection pair of multiple barrier characteristicsThe contextual data and corresponding multiple classifier parameters answered simultaneously carry out operation, to obtain more classification data, then by classification dataCompared with a floating threshold, the corresponding obstacle species of floating threshold, and the institute's frame for exporting corresponding floating threshold selects barrierThe position of image.

Below by specific embodiment elaborate, when be easier to understand the purpose of the present invention, technology contents, feature and itsThe effect of reached.

Detailed description of the invention

Fig. 1 is block schematic diagram of the invention；

Fig. 2 is step flow chart of the invention；

Fig. 3 is that the multiple frame of parallelization in the present invention selects barrier image schematic diagram；

Fig. 4 A to Fig. 4 D is the flow diagram of barrier in detecting image data of the present invention.

Description of symbols: 10- reservoir；102- scenario parameters；104- classifier parameters；12- video capture device；122-Image data；124,124a, 124b, 124c, 124d- barrier image；14- image processor；142,142a,142b,142c,142d- image process unit；144- adjusts dimension cells；16- display.

Specific embodiment

To make anti-collision system (AEB) that can quickly judge braking time, the present invention provides one kind to have parallelization frameThe adaptability object sorter and its method of structure and pipelined operation technology, come solve the prior art can not detecting real-time lackPoint.

It is as shown in Figure 1 block schematic diagram of the invention.Adaptability object sorter is built into an anti-collision system, is fittedAnswering property object classification device includes a reservoir 10, at least a video capture device 12 and an image processor 14, image processor14 are electrically connected reservoir 10 and video capture device 12.Wherein, reservoir 10 can be synchronous dynamic random-access reservoir(Synchronous Dynamic Random Access Memory, SDRAM), internal reservoir has more scenario parameters 102And more classifier parameters 104, image processor 14 include multiple image process units 142 and multiple adjustment dimension cells(resize unit) 144, multiple image process unit 142 are electrically connected multiple adjustment dimension cells 144.Image captureAfter device 12 captures an image data 122, frame selects multiple barrier images 124 in image data 122, and according to each barrier124 range of object image is hindered to select selection range appropriate.Since video capture device 12 has far and near pick-up image ability, barrier124 range of image is also different therewith, therefore each barrier image 124 for being selected institute's frame using multiple adjustment dimension cells 144Range is adjusted to selection range appropriate, can so reduce the data volume of subsequent images processing.In order to accelerate image processing to imitateRate, center select several barrier images 124 and are just corresponding with the several progress of image process units 142 parallel processing, Yi JiyingPicture processing unit 142 and barrier image 124 are one-to-one parallel processing, therefore multiple image process units 142 are correspondingCalculate more barrier characteristics of multiple barrier images.Corresponding field is chosen further according to multiple barrier characteristicsScape parameter 102 and corresponding multiple classifier parameters 104 simultaneously carry out operation, to obtain more classification data, then by classification dataCompared with a floating threshold, the corresponding obstacle species of floating threshold, and the institute's frame for exporting corresponding floating threshold selects barrierThe position of image 124 is in a display 16.Using several 124 methods of barrier image of parallel processing, can speed up at imageEfficiency is managed, and then simplifies the complexity of image processing and reduces time-consuming.

Reach detecting real-time effect to be further understood that the present invention such as how parallel shelf structure and pipelined operation mode,As shown in Figure 1, Figure 2 and Figure 3, Fig. 2 is step flow chart of the invention, and Fig. 3 is that the multiple frame of parallelization of the invention selects barrierImage schematic diagram.Firstly, capturing an at least image data 122 such as step S10, illustrating by taking an image data as an example herein.It canAccording to dynamic interest section of driving a vehicle at present, dynamically adjust the region scanned, avoid scanning the imagery zone for not needing operation withOperand is reduced, such as step S12, an interesting region (ROI) is selected from image data, then select from interesting region centerMultiple barrier images in image data 122, multiple barrier images are vehicle, pedestrian, animal, electric pole, Lu Shu, roadblockOr above combination, as shown in figure 3, multiple barrier images include barrier image 124a, barrier image 124b, obstacleObject image 124c and barrier image 124d is respectively corresponded every to adjust such as step S14 using multiple adjustment dimension cells 144The selection range size of one barrier image 124a, 124b, 124c, 124d is chosen to be each 124 range of barrier image suitableWhen selection range, due to the difference of obstacle species, length and the shadow presented in view of shooting focal length is far and nearAs difference in size, selection range size is adjusted for different barrier images using adjustment dimension cells 144, is able to ascend shadowAs processing speed.Such as step S16, more barriers for calculating multiple barrier images are respectively corresponded using multiple image process unitsHinder object characteristic, as image process unit 142a calculates the barrier characteristic of barrier image 124a, image processingUnit 142b calculates the barrier characteristic of barrier image 124b, and image process unit 142c calculates barrier imageThe barrier characteristic of 124c, image process unit 142d calculate the barrier characteristic of barrier image 124d,In, image process unit 142a, 142b, 142c, 142d be barrier image 124a, 124b, 124c corresponding to parallel processing,124d, and corresponding scenario parameters 102 and corresponding multiple classifier parameters 104 are chosen simultaneously according to multiple barrier characteristicsOperation is carried out, to obtain more classification data.Wherein, scenario parameters 102 are overexposure scene, night scenes, on the sunny side scene or HuangConfused scene, and elastic according to different scenes can adjust, classifier parameters 104 include more direction ladders of different obstacle speciesCharacteristic parameter is spent, and can be adjusted in time according to obstacle species.Whether last such as step S18, match stop data are greater than oneFloating threshold, the corresponding obstacle species of floating threshold determine the corresponding barrier of classification data if so, thening follow the steps S20Type, and export the position that institute's frame selects barrier image；If it is not, thening follow the steps 22, then it is any to determine that classification data does not correspond toOne obstacle species, wherein obstacle species are car body, pedestrian or roadblock.Since the present invention can sentence according to the image capturedDisconnected corresponding scene out, it is higher can to calculate accuracy according to various different scenes come elastic adjustment classifier parameters 104Barrier image 124 and its type, to solve the problems, such as the erroneous judgement of existing detecting system, and in multiple image process units 142It is built in the anti-collision system of vehicle.

To further illustrate how image process unit 142 of the invention calculates the barrier characteristic of barrier imageAccording to, and how corresponding scenario parameters and corresponding classifier parameters to be chosen according to barrier characteristic and carry out operation,To obtain classification data, illustrate by taking pedestrian's detecting image as an example in this barrier image.As shown in Fig. 1 and Fig. 4 A to Fig. 4 D, figure4A to Fig. 4 D is the flow diagram of barrier in detecting image data of the present invention.Here, the present invention is that use direction gradient is calculatedMethod (Histogram of Oriented Gradient, HOG) calculates barrier characteristic value, predominantly counts the ladder of whole imageThe foundation of intensity and directional information as subsequent classification is spent, use direction gradient algorithm increases intensity for the edge of barrierAnd there is higher tolerance to barrier deformability.In addition, the more classifier parameters that the present invention stores are from supporting vectorMachine classifier (Support Vector Machine, SVM), to be corresponded to the characteristic parameter of barrier using inner product of vectorsThe separating degree of barrier characteristic value is maximum by one hyperplane.Specifically, support vector machine classifier is using hundreds of or countsThe pedestrian sample of a thousand sheets after calculating barrier characteristic value via direction gradient algorithm, then is inputted as pedestrian's image databaseTrained off-line is carried out to support vector machine classifier, finally using more classifier parameters of training result as subsequent obstacle speciesThe classification foundation of class.

As shown in Figure 4 A, an image data 122 is captured, and frame selects multiple barrier images in image data 122124a, barrier image 124b, barrier image 124c, then as shown in Figure 4 B, each 142 utilization orientation of image process unit ladderDegree algorithm first calculates the marginal value of each barrier image, and each barrier image 124 after frame is selected is divided into multiple smallBlock (block), then calculate the direction of each pixel in each block of cells and edge parameter value and add up, it is each to obtainNine feature vectors in block of cells.For example, detect barrier image 124 be pedestrian when, 124 size of barrier imageIt is 64 (n) × 128 (m), each of barrier image 124 pixel is subjected to edge detection (Edge Detection), it canThe edge direction and edge strength of each pixel are obtained, it is 16 × 16 and portion that barrier image 124, which is next divided into size,The operation for dividing the unit (cell) to overlap to carry out direction gradient algorithm, and 8 lattice of displacement every time, here, in order to reduce operation time,Pipeline design method can be utilized, data flow is pre-stored in a fritter memory (SRAM), to reduce the data-moving time,In the assisted instruction period, wherein memory is electrically connected image processor 14.Since edge direction difference 180 degree can be considered same sideTo, therefore each unit is divided into nine feature vectors, that is, each 16 × 16 blocks in 0~180 degree according to edge directionNine directions can be corresponded to, as shown in Figure 4 C, the total amount of data in this stage is a barrier characteristic of 9 × (n/8) × (m/8)According to.Wherein all pixels do ballot statistics to the direction character vector belonging to it respectively in each unit, and the poll thrown is pictureThe information of the edge strength of element, this nine directions can be represented with the vector of nine dimensions, that is, nine feature vectors.By four phasesAdjacent unit is considered as a block, can mutually overlap between different blocks, and block describes obstacle with the feature vector of its interior 4 unitThe local edge information of 124 position of object image.Parallel calculation is utilized herein, four units are calculated with the time, and willIt is put into memory to be counted, to accelerate the whole instruction cycle.It is finally represented with 36 dimensional vectors, 36 dimensional vectors are through normalization(Normalize) make vector length 1, by 36 dimensional vectors of all 7x15 blocks combine it is available 3780 dimension toAmount, this vector contain the whole information with part of pedestrian, that is, the barrier characteristic of barrier image 124.

It connects, every barrier characteristic is multiplied by corresponding classifier parameters, such as barrier characteristic 9 ×(n/8) × (m/8) is multiplied by classifier parameters 9 × (n/8) × (m/8), and adds up and obtain an end value, this end value frame thusThe score of favored area represents this frame favored area with detected barrier as soon as such score is greater than floating threshold, it is on the contrary thenClear, as shown in Figure 4 D, herein for detecting pedestrian, therefore, the score of frame favored area is greater than one floating threshold, that is, tableShow that frame favored area has detected barrier for pedestrian, and pedestrian's image of display box favored area, such as shows barrier image124b, barrier image 124c；Conversely, the score of frame favored area is less than a floating threshold, then frame favored area, which has, is detectedBarrier non-pedestrian, thus it is for example possible to be vehicle, roadblock or other barriers or without any barrier.Not due to obstacle speciesTogether, classifier parameters can also change therewith with floating threshold, therefore can adjust classifier parameters in real time, more accurately be detectd with reachingEfficiency is surveyed, existing generate because scenario parameters and sorting parameter can not be adaptively adjusted is no longer limited to and detects asking for mistakeTopic.That is, image process unit 142 can choose corresponding scenario parameters and corresponding classification according to barrier characteristicDevice parameter simultaneously carries out operation, to obtain classification data, reinforces the identification precision of this device.In conclusion due to the prior artObstruction detection and sort out obstacle species must in image one by one search detect barrier, then sequentially detect operation atObstacle species are managed and sort out, this image operand is quite big and time-consuming, is not easy to reach the efficiency of real-time judge.And it is of the inventionIn order to enable anti-collision system quickly accurately to judge brake opportunity, therefore devise the suitable of the parallel framework of tool and pipelined operation frameworkAnswering property object classification device can speed up statistical vector using the multiple barrier images of one-to-one image processing parallel calculationTreatment effeciency, the higher barrier data of accuracy are provided and give anti-collision system, are avoided that traffic accident, and then simplify shadowAs the complexity and time-consuming problem of processing.

In addition, since current used classifier parameters are the good parameters of precondition, and be embedded in anti-collision system, whenWhen vehicle encounters different scenes in outdoor traveling, such as dusk, on the sunny side, night or overexposure scene, it is appropriate because that can not adjustScenario parameters also can not adjust classifier parameters because scene is different, erroneous judgement and the decline of detecting rate etc. is be easy to cause to lackPoint.And the present invention can judge corresponding scene according to the image captured, it can be according to various different scenes come elastic adjustment pointClass device parameter can calculate the higher barrier image of accuracy and its kind because having the scene judgement of adaptabilityClass, to solve the problems, such as existing detecting system erroneous judgement.Still further, the present invention can according to the remote deep of the image captured andThe size of barrier image, frame select barrier image capturing range appropriate and carry out image processing again, are able to ascend image processing effectRate, so solve the problems, such as that current image capturing range is big and caused by operand it is huge.

The foregoing is merely illustrative of the preferred embodiments of the present invention, is not used to limit the scope of implementation of the present invention.Therefore i.e.Equivalent change or modification carried out by all features according to the scope of claims of the present invention and spirit, should be included in of the inventionIn protection scope.

Claims (12)

Translated fromChinese

1.一种具平行架构的适应性物体分类方法，该方法包含储存多笔场景参数及多笔分类器参数的步骤，其特征在于，该适应性物体分类方法包括下列步骤：1. an adaptive object classification method with a parallel structure, the method comprising the steps of storing multiple scene parameters and multiple classifier parameters, it is characterized in that, this adaptive object classification method comprises the following steps:撷取至少一影像数据；capturing at least one image data;框选出该影像数据中的多个障碍物影像，且根据每一该障碍物影像范围选定适当的框选范围；frame selection of a plurality of obstacle images in the image data, and select an appropriate frame selection range according to each obstacle image range;利用多个影像处理单元分别对应计算出该多个障碍物影像的多笔障碍物特征数据，并根据该多个障碍物特征数据选取对应的该场景参数及对应的该多个分类器参数并进行运算，以取得多笔分类数据，每一该分类数据由每一该障碍物特征数据乘上相对应的该分类器参数，并加总而取得；以及Using a plurality of image processing units to calculate a plurality of pieces of obstacle feature data corresponding to the plurality of obstacle images, and select the corresponding scene parameters and the corresponding plurality of classifier parameters according to the plurality of obstacle feature data, and perform an operation to obtain a plurality of classification data, each of the classification data is obtained by multiplying each of the obstacle feature data by the corresponding classifier parameter, and summing them up; and比较该多个分类数据是否大于一浮动阀值，该浮动阀值对应一障碍物种类并根据该障碍物种类及该多个场景参数不同而对应的阀值，若是，则判定该分类数据对应的该障碍物种类，并输出所框选该障碍物影像的位置，若否，则判定分类数据不对应任何一个该障碍物种类。Compare whether the plurality of classification data is greater than a floating threshold, the floating threshold corresponds to an obstacle type and the corresponding threshold according to the obstacle type and the plurality of scene parameters, if yes, then determine the corresponding classification data. the obstacle type, and output the position of the framed image of the obstacle; if not, it is determined that the classification data does not correspond to any obstacle type.2.根据权利要求1所述的具平行架构的适应性物体分类方法，其特征在于，于框选该多个障碍物影像的步骤之前，进一步包括以下步骤：从该影像数据中选定一有兴趣区域，再从该有兴趣区域中框选出该影像数据中的该多个障碍物影像。2 . The adaptive object classification method with a parallel structure according to claim 1 , wherein before the step of frame-selecting the plurality of obstacle images, the method further comprises the following step: selecting an object from the image data. 3 . area of interest, and then frame and select the plurality of obstacle images in the image data from the area of interest.3.根据权利要求2所述的具平行架构的适应性物体分类方法，其特征在于，于框选该多个障碍物影像步骤之后，进一步包括以下步骤：调整每一该障碍物影像的该框选范围大小。3 . The adaptive object classification method with a parallel structure as claimed in claim 2 , wherein after the step of framing the plurality of obstacle images, the method further comprises the following step: adjusting the frame of each obstacle image. 4 . Select the range size.4.根据权利要求1所述的具平行架构的适应性物体分类方法，其特征在于，每一该影像处理单元利用一方向梯度算法将每一该障碍物影像划分为多个小区块，并统计每一该小区块内的九个特征向量，获得该障碍物影像的该障碍物特征数据。4 . The adaptive object classification method with a parallel structure according to claim 1 , wherein each of the image processing units divides each of the obstacle images into a plurality of small blocks by using a directional gradient algorithm, and counts the For each of the nine feature vectors in the small block, the obstacle feature data of the obstacle image is obtained.5.根据权利要求4所述的具平行架构的适应性物体分类方法，其特征在于，该方向梯度算法包括以下步骤：先计算每一该障碍物影像的边缘值，再计算每一该小区块中的每一像素的方向及边缘参数值并加总，以取得该九个特征向量。5 . The adaptive object classification method with parallel architecture as claimed in claim 4 , wherein the directional gradient algorithm comprises the following steps: firstly calculating the edge value of each of the obstacle images, and then calculating each of the small blocks. 6 . The direction and edge parameter values of each pixel in are summed to obtain the nine feature vectors.6.根据权利要求1所述的具平行架构的适应性物体分类方法，其特征在于，该多个场景参数包含过曝场景、夜间场景、向阳场景或黄昏场景；该多个分类器参数包括不同障碍物种类的多笔方向梯度特征参数。6 . The adaptive object classification method with a parallel structure according to claim 1 , wherein the plurality of scene parameters include an overexposure scene, a night scene, a sunny scene or a twilight scene; the plurality of classifier parameters include different Multi-pen directional gradient feature parameters for obstacle types.7.一种具平行架构的适应性物体分类装置，该装置储存有多笔场景参数及多笔分类器参数，其特征在于，该适应性物体分类装置包括：7. An adaptive object classification device with a parallel structure, the device stores multiple scene parameters and multiple classifier parameters, wherein the adaptive object classification device comprises:至少一影像撷取器，用于撷取至少一影像数据，并框选出该影像数据中的多个障碍物影像，且根据每一该障碍物影像范围选定适当的框选范围；以及at least one image capture device for capturing at least one image data, and frame selection of a plurality of obstacle images in the image data, and select an appropriate frame selection range according to each obstacle image range; and一影像处理器，电性连接该影像撷取器，该影像处理器包含多个影像处理单元，多个影像处理单元用于分别对应计算出该多个障碍物影像的多笔障碍物特征数据，根据该多个障碍物特征数据选取对应的场景数据及对应的分类器参数并进行运算，以取得多笔分类数据，每一该分类数据由每一该障碍物特征数据乘上相对应的该分类器参数，并加总而取得，再将该多个分类数据与一浮动阀值比较，该浮动阀值对应一障碍物种类，根据该障碍物种类及该多个场景参数不同而对应的阀值，并输出相对应该浮动阀值的所框选该障碍物影像的位置。an image processor, electrically connected to the image capture device, the image processor includes a plurality of image processing units, and the plurality of image processing units are used to calculate a plurality of pieces of obstacle feature data corresponding to the plurality of obstacle images, respectively, According to the plurality of obstacle feature data, corresponding scene data and corresponding classifier parameters are selected and calculated to obtain multiple pieces of classification data, each of which is multiplied by each of the obstacle feature data by the corresponding classification Then, the plurality of classification data are compared with a floating threshold value, the floating threshold value corresponds to an obstacle type, and the threshold value corresponding to the obstacle type and the plurality of scene parameters is different. , and output the position of the framed obstacle image relative to the floating threshold.8.根据权利要求7所述的具平行架构的适应性物体分类装置，其特征在于，该影像撷取器能够从该影像数据中选定一有兴趣区域，再从该有兴趣区域中框选出该影像数据中的该多个障碍物影像。8 . The adaptive object classification device with a parallel structure as claimed in claim 7 , wherein the image capture device is capable of selecting an area of interest from the image data, and then making a frame selection from the area of interest. 9 . outputting the plurality of obstacle images in the image data.9.根据权利要求8所述的具平行架构的适应性物体分类装置，其特征在于，该影像处理器包含多个调整尺寸单元，多个调整尺寸单元分别电性连接该多个影像处理单元，多个调整尺寸单元根据该影像撷取器撷取该影像数据的远近，利用每一该调整尺寸单元对应调整每一该障碍物影像的该框选范围大小。9 . The adaptive object classification device with a parallel structure according to claim 8 , wherein the image processor comprises a plurality of resizing units, and the plurality of resizing units are respectively electrically connected to the plurality of image processing units, 10 . A plurality of resizing units are used to adjust the size of the frame selection range of each obstacle image by using each of the resizing units according to the distance of the image data captured by the image capture unit.10.根据权利要求7所述的具平行架构的适应性物体分类装置，其特征在于，每一该影像处理单元利用一方向梯度算法将每一该障碍物影像划分为多个小区块，并统计每一该小区块内的九个特征向量，获得该障碍物影像的该障碍物特征数据。10 . The adaptive object classification device with a parallel structure according to claim 7 , wherein each of the image processing units divides each of the obstacle images into a plurality of small blocks using a directional gradient algorithm, and counts the For each of the nine feature vectors in the small block, the obstacle feature data of the obstacle image is obtained.11.根据权利要求7所述的具平行架构的适应性物体分类装置，其特征在于，该多个场景参数包含过曝场景、夜间场景、向阳场景或黄昏场景；该多个分类器参数包括不同该障碍物种类的多笔方向梯度特征参数。11 . The adaptive object classification device with a parallel structure according to claim 7 , wherein the plurality of scene parameters include an overexposure scene, a night scene, a sunny scene or a twilight scene; the plurality of classifier parameters include different The multi-pen directional gradient feature parameters of the obstacle type.12.根据权利要求7所述的具平行架构的适应性物体分类装置，其特征在于，该影像撷取器及该影像处理器内建于一车辆的防撞系统中。12 . The adaptive object classification device with a parallel structure as claimed in claim 7 , wherein the image capture device and the image processor are built in a collision avoidance system of a vehicle. 13 .