JP2010186433A - Target detector - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a target detector for suppressing the spread of a target pixel, improving the detection accuracy of a target and detecting a plurality of proximity targets. <P>SOLUTION: The target detector includes: a sensor part 1 for observing a target; an observation image acquisition part 2 for acquiring an observed image; a likelihood ratio calculation part 3 for calculating the likelihood ratio of each pixel in the observation image; a transition probability calculation part 4 for calculating the transition probability of each pixel in the observation image; an integration processing part 5 for performing the integration processing of the likelihood ratio, the transition probability and a track score which is the processing result of one frame before; a redundant allocation extraction part 6 for extracting a part where the plurality of pixels of a present frame image are made to correspond to the same pixel of the image of one frame before and redundant allocation is generated from an integration processing result; a redundant allocation dissolving part 7 for dissolving the extracted redundant allocation; an end determination part 8 for determining the end of the integration processing; and a target detection part 9 for detecting the pixel of a high track score as the target from a finally obtained processing result. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a target detection device that detects a target (particularly a minute target) in a low SNR (Signal to Noise Ratio) image acquired by an observation sensor.

  As a conventional technique for detecting a target in a low SNR image, for example, there is Non-Patent Document 1. This is because the target in the low SNR image is buried in noise and it is difficult to detect the target from a single image. Therefore, the target detection is performed by integrating the images of a plurality of frames taken continuously. How to do it.

  In Non-Patent Document 1, using the likelihood ratio of each pixel (representing the probability that the pixel is the target) and the transition probability of each pixel from the previous frame to the current frame based on the motion model defined in advance, Target detection is performed based on the track score defined by equation (1). FIG. 13 shows a block diagram of processing shown in Non-Patent Document 1.

  The track score shown in Equation (1) is an integration process that takes into account the pixel motion and likelihood ratio based on the target motion model, so the position where the target exists can be calculated by consecutively calculating multiple frames. The track score is greatly increased compared to other pixels, and it is easy to detect a minute target in a low SNR image.

  Moreover, there exists patent document 1 as invention which improved the method of nonpatent literature 1. FIG. Patent Document 1 is an invention aimed at detecting a target with a very low SNR of 0 or less.

J. Arnold, S. Shaw, H. Pasternack “Efficient Target Tracking Using Dynamic Programming”, IEEE Trans. Aerospace and Electronic Systems, Vol. 29, No. 1, pp. 44-56 (1993).

JP 2008-275339 A

  However, in the technique shown in Non-Patent Document 1, each pixel of frame k independently selects a pixel of frame k−1 that maximizes its track score, and thus a plurality of pixels of frame k A situation occurs in which the same pixel of k−1 is selected. This situation is called duplicate assignment. FIG. 14 schematically shows overlapping assignment. FIG. 14 shows a situation in which the pixel A, the pixel B, and the pixel C in the frame k select the pixel X in the frame k−1. In the method of Non-Patent Document 1, such a situation occurs because the pixel of the frame k−1 that maximizes the track score when viewed from each pixel of the frame k is selected. Conversely, one pixel of frame k does not correspond to a plurality of pixels of frame k-1.

  The situation as shown in FIG. 14 indicates that the target of one pixel in the k-1 frame spreads to a plurality of pixels in the k frame, and the spread gradually increases as the integration process proceeds. For this reason, in the processing result, the target pixels, which should be about several pixels, have a large spread, making it difficult to accurately detect the target position, and it becomes impossible to determine the target when there are multiple adjacent targets. There is. FIG. 15 shows the result of integration processing by the method shown in Non-Patent Document 1.

  Further, the invention of Patent Document 1 does not show a solution to the above problem.

  The present invention has been made in order to solve the above-described problem. In consideration of the target motion model, the overlap allocation is eliminated and the spread of the target pixel is suppressed so as to minimize the decrease in the track score. Thus, an object of the present invention is to obtain a target detection device that can improve the detection accuracy of a target and enable detection of a plurality of adjacent targets.

  A target detection apparatus according to the present invention is a target detection apparatus that enables target detection even in a low SNR environment by integrating a plurality of continuously captured observation images, and is a sensor that performs target observation An observation image acquisition unit that acquires an image observed by the sensor unit, a likelihood ratio calculation unit that calculates a likelihood ratio of each pixel in the acquired observation image, and an acquired observation according to a target assumed motion model The transition probability calculation unit that calculates the transition probability of each pixel in the image, the likelihood ratio obtained by the likelihood ratio calculation unit, the transition probability obtained by the transition probability calculation unit, and the processing result of one frame before From the integration processing unit that performs integration processing with the track score, and the integration processing result by the integration processing unit, a plurality of pixels of the current frame image are associated with the same pixel of the image one frame before, and overlapping allocation occurs. Where A duplicate allocation extraction unit to extract, a duplicate allocation elimination unit that eliminates the duplicate allocation extracted by the duplicate allocation extraction unit, an end determination unit that determines the end of integration processing based on the number of integration frames or integration processing time, A target detection unit that detects a pixel having a track score higher than the finally obtained processing result as a target, and the duplication allocation extraction unit includes a plurality of current frame images for the same pixel of the image one frame before Extracting a non-assigned pixel of an image one frame before that is not associated with a pixel where a duplicate assignment has been assigned and a pixel of the current frame image, and the duplicate assignment canceling unit Reassign between pixels so that the pixel of the current frame image with the maximum track score remains for the allocation occurrence location, and the other pixels and the unallocated pixels of the image one frame before correspond one-to-one. When the pixel of the current frame image repeatedly allocating no reallocation performed has not been resolved et exists and replaces the track score to a value of only likelihood ratio considers the pixel and the initial detection pixels.

  According to the present invention, in consideration of the target motion model, the target detection accuracy is improved by eliminating the duplicate assignment and suppressing the spread of the target pixel so as to minimize the decrease in the track score. Multiple targets can be detected.

It is a block diagram which shows the structure of the target detection apparatus in Embodiment 1 of this invention. It is a figure which shows the example in case there exist two duplication generation | occurrence | production locations. It is the figure which showed the outline | summary of the duplicate allocation cancellation process. It is a figure explaining the extraction process of a transition source range. It is a block diagram which shows the structure of the target detection apparatus in Embodiment 2 of this invention. It is a figure explaining an allocation fixed pixel determination process. It is a block diagram which shows the structure of the target detection apparatus in Embodiment 3 of this invention. FIG. 10 is a diagram used for explaining operations of a connected region extracting unit 13 and a connected region reducing unit 14. It is the figure which showed the connection relation of the pixel. It is the figure which showed the example of the connection area | region. FIG. 10 is a diagram used for explaining the operation of the connected region reduction unit 14. It is a block diagram which shows the structure of the target detection apparatus in Embodiment 4 of this invention. Non-Patent Document 1 (J. Arnold, S. Shaw, H. Pasternack “Efficient Target Tracking Using Dynamic Programming”, IEEE Trans. Aerospace and Electronic Systems, Vol. 29, No. 1, pp. 44-56 (1993)) It is a block diagram which shows the content of the process shown by FIG. It is the figure which showed the duplication allocation typically. It is a figure which shows the integration process result by the method shown by the nonpatent literature 1.

  The best mode for carrying out the present invention will be described below with reference to the drawings.

Embodiment 1 FIG.
1 is a block diagram showing a configuration of a target detection apparatus according to Embodiment 1 of the present invention. The target detection apparatus shown in FIG. 1 is a target detection apparatus that enables target detection even in a low SNR environment by integrating a plurality of continuously captured images. A space within a certain range is observed and an observation image is generated. The observation image acquisition unit 2 acquires the observation image generated by the sensor unit 1, and the likelihood ratio calculation unit 3 calculates the likelihood ratio of each pixel in the image acquired by the observation image acquisition unit 2.

  Based on the assumed target motion model, the transition probability calculation unit 4 has a probability that the target moves from the pixel (m, n) of the image acquired one frame before to the pixel (i, j) of the image of the current frame ( Called transition probability). The integration processing unit 5 calculates the likelihood ratio obtained by the likelihood ratio calculation unit 3, the transition probability obtained by the transition probability calculation unit 4, the track score of the image one frame before (the processing result for each frame is referred to as a track score). To obtain the track score of the image of the current frame, and the correspondence between each pixel of the current frame image and the pixel of the image one frame before that maximizes the track score of each pixel of the current frame image I do.

  The duplication allocation extraction unit 6 extracts the location where the duplication allocation as shown in FIG. 14 has occurred, based on the result of association between each pixel of the current frame image and the previous frame image in the integration processing unit 5. To do. The duplicate allocation cancellation unit 7 eliminates the duplicate allocation for the duplicate allocation occurrence location extracted by the duplicate allocation extraction unit 6 and sets each pixel of the image of the current frame and each pixel of the image of the previous frame as much as possible. Recalculate the track score of the pixel of image k corresponding to the pair 1 and eliminating the duplicate assignment.

  The end determination unit 8 determines whether or not the total number of frames or the integration processing time has reached the upper limit value. If the upper limit value has not been reached, the process is returned to the observation image acquisition unit 2, and if the upper limit value has been reached. The addition process is terminated. The target detection unit 9 detects a pixel having a maximum track score after the addition process as a target.

  Next, a series of operations of the target detection apparatus having the configuration of FIG. 1 will be described. First, the sensor unit 1 periodically observes the determined space and generates an observation image. The observation data acquisition unit is called a frame, and the observation image of frame k is hereinafter referred to as image k. Images with adjacent frame numbers (for example, image k-1 and image k) are continuously observed images.

Next, the observation image acquisition unit 2 acquires an observation image from the sensor unit 1 and inputs the observation image to the likelihood ratio calculation unit 3 and the transition probability calculation unit 4. Based on this, the likelihood ratio calculation unit 3 calculates the likelihood ratio of each pixel of the observation image. The likelihood ratio is an index indicating the probability that each pixel is a target. The amplitude value (luminance value) of the pixel (i, j) of the image k is z ^k _ij and the pixel (i, j) is the target. When the probability density function in a certain case is p _t (x) and the probability density function in the case where the pixel (i, j) is noise is p _n (x), the likelihood ratio of the pixel (i, j) of the image k L is defined by equation (2).

  Next, the transition probability calculation unit 4 uses the predicted position of the target obtained using a tracking filter such as an α-β filter based on the assumed target motion model, for example, the pixel (m, The probability that the target moves to the pixel (i, j) of the image k from n) is calculated (this is called the transition probability).

Since the assumed target motion model differs depending on the application destination, the definition of the transition probability differs depending on the application destination and cannot be determined uniquely, but an example of the definition expression is shown in Expression (3). Equation (3), the image k-1 of the pixel (m, n) the movement predicted position of the target from the image k in the case of ^{_{(I k m, n, J}} k m, n) and the image k-1 Is a transition probability from pixel (m, n) to pixel (i, j) of image k, C is a coefficient, and σ is a standard deviation of the transition probability. Both C and σ are values given in advance. Further, g ^k _{i, j} represents a pixel (i, j) of the image k, and g ^k−1 _{m, n} represents a pixel (m, n) of the image k−1.

  Next, the integration processing unit 5 integrates the likelihood ratio calculated by the likelihood ratio calculation 3, the transition probability obtained by the transition probability calculation unit 4, and the track score of the image k−1 according to the equation (4). A track score of k is obtained, and each pixel of the image k is associated with a pixel of the image k−1 that maximizes the track score of each pixel of the image k.

The process shown in Expression (4) follows the following procedure.
step 1. The natural logarithm of the transition probability from the pixel (m, n) to the pixel (i, j) of the image k out of the pixel (m, n) of the image k−1 and the pixel ( The pixel having the maximum track score of m, n) is selected and associated with the pixel (i, j) of the image k. For example, assuming that the pixel (4, 3) of the image k-1 and the pixel (6, 5) of the image k are associated with each other, if the target exists in the pixel (4, 3) of the image k-1, the image k-1 Represents that the target has been moved to the pixel (6, 5).

  step 2. The total of the maximum value obtained in step 1 and the natural logarithm of the likelihood ratio of the pixel (i, j) of the image k is defined as the track score of the pixel (i, j) of the image k.

  Next, the duplicate allocation extraction unit 6 extracts a duplicate allocation occurrence location from the result of the association between the pixel of the image k−1 and the pixel of the image k in the integration processing unit 5. The processing procedure in the duplicate allocation extraction unit 6 is shown below.

step 1. Extraction of duplicate groups Extraction locations where duplicate allocation occurs as shown in FIG. 14 are extracted. If there are multiple occurrence locations, group numbers are assigned. FIG. 2 shows an example in which there are two overlapping occurrence locations. In this case, a group of pixels {X, A, B, C} is an overlap group 1, and a group of pixels {W, D, E, F, G} is an overlap group 2.

step 2. Extraction of assigned pixels There are also pixels that are associated one-to-one between image k-1 and image k. Among these, the pixel on the image k-1 side is extracted as “allocated pixel”.

step 3. Extraction of unassigned pixels Image k-1 and image k are images observed by the same sensor in a series of processing, and the number of pixels can be considered to be the same. Therefore, in a situation where overlapping assignment occurs between the image k-1 and the image k, the image k-1 includes pixels that do not correspond to the pixels of the image k. Such a pixel is referred to as an “unallocated pixel”, and an unallocated pixel is extracted from the image k−1.

  Next, the duplicate allocation elimination unit 7 eliminates the duplicate assignment of each duplicate assignment occurrence location extracted by the duplicate assignment extraction unit 6, and corresponds to each pixel of the image k-1 and the image k as much as possible. Let Hereinafter, the operation of the duplicate allocation elimination unit 7 will be described.

  FIG. 3 is a diagram showing an outline of the duplicate allocation elimination process. In the image k-1 in FIG. 3, pixels filled in black indicate assigned pixels, and white pixels indicate unassigned pixels. Under such circumstances, the pixel X of the image k-1 and the pixels A, B, and C of the image k are in an overlapping allocation relationship (upper part of FIG. 3). For the pixels A, B, and C having the overlapping relationship, the allocation destination of the remaining pixels is selected from the unallocated pixels of the image k−1 while leaving one pixel. In the example of FIG. 3, the pixel A is reassigned to the pixel Y of the image k-1, and the pixel C is reassigned to the pixel Z of the image k-1.

The detailed operation of the duplicate allocation elimination unit 7 will be described below.
step 1. Prioritizing duplicate groups Priorities are assigned to the duplicate groups extracted by the duplicate assignment extraction unit 6 in descending order of the track scores of the pixels of the image k-1 included in each duplicate group. Referring to FIG. 2 as an example, the track score of the pixel X that is the pixel of the image k-1 of the overlap group 1 is compared with the track score of the pixel W that is the pixel of the image k-1 of the overlap group 2, and If the track score value is high, the priority of the overlap group 1 is increased, and conversely, if the track score value of the pixel W is high, the priority of the overlap group 2 is increased. Duplicate assignment elimination shown in the next step and subsequent steps is performed on all duplicate groups in descending order of priority.

  In the above description, the value of the track score of the pixel on the image k-1 side is used as the reference for prioritization, but (1) the maximum value of the track score of the pixel of the image k included in each overlapping group, (2) The number of pixels of the image k included in each overlapping group may be used as a reference. The example in the case of (1) will be described with reference to FIG. 2. The track score of the pixel B is the largest among the pixels of the image k belonging to the overlap group 1, and the pixel F is the pixel of the image k belonging to the overlap group 2. If the track score of pixel B is the largest, the priority of overlap group 1 is higher if the track score of pixel B is greater than the track score of pixel F, and conversely, if the track score of pixel F is greater, Increase priority. Further, the case of (2) will be described with reference to FIG. 2. Since the number of pixels of the image k belonging to the overlap group 1 is 3 and the number of pixels of the image k belonging to the overlap group 2 is 4, the priority order of the overlap group 2 is set. Make it high.

  Further, the above three criteria are combined, for example, prioritization is first performed based on the track score of the pixels of the image k-1 included in each overlap group. If the priorities are the same, the image k included in each overlap group The prioritization may be performed by the maximum track score of the pixels, and when the priorities are the same, the prioritization may be performed by the number of pixels of the image k included in each overlapping group.

step 2. Prioritize the pixels in the overlap group Prioritize the pixels of image k in the same overlap group in descending order of the track score, and exclude the pixel with the highest priority from the duplicate allocation elimination process (ie , The current allocation is maintained), and the allocation cancellation processing shown in the next step and subsequent steps is performed on all the pixels having the second and higher priorities in descending order of priority. Note that the pixels with the second and subsequent priorities are referred to as reallocation target pixels.

  In the above description, the track score of the pixel of the image k is used as a priority ranking criterion. However, the priority ranking may be performed in descending order of the likelihood ratio of each pixel of the image k.

step 3. Extraction of transition source range Considering the assumed motion model, the pixel range of image k-1 that can transition to the Nth priority (N ≧ 2) reassignment target pixel (transition source) (Referred to as a range). Furthermore, unallocated pixels of the image k−1 included in the transition source range are extracted as transitionable pixels.

  The process of step 3 will be described with reference to FIG. The upper part of FIG. 4 shows the occurrence of overlapping assignment, and the pixels A, B, and C of the image k are assigned to the pixel X of the image k−1. At this time, when the track score of B is the highest among the pixels A, B, and C, the assignment of the pixel B to the pixel X is left as it is. Accordingly, the reassignment targets are the pixel A and the pixel C, and FIG.

  In step 3, first, the transition source range of the image k-1 that can transition to the pixel A is extracted. The transition source range is extracted by calculating the maximum number of pixels R that the target can move in one frame based on the assumed target motion model, and centering on the same position as the pixel A of the image k−1 (2R + 1). ) × (2R + 1) pixel area is set as the transition source range (see the lower left in FIG. 4). Next, unassigned pixels included in the transition source range are extracted as transitionable pixels (see the lower center of FIG. 4).

step 4. Reallocate
From the transitionable pixels extracted in step 3, select the pixel having the maximum track score of the allocation target pixel and associate it with the allocation target pixel. However, when a transitionable pixel is not extracted in step 3, the allocation target pixel is determined as a non-reallocable pixel.

  In the example on the right side of the lower part of FIG. 4, the pixel Y is associated with the pixel A, assuming that the pixel Y among the transitionable pixels extracted in step 3 maximizes the track score of the pixel A.

step 5. Non-reallocable pixel processing (initial detection)
After executing the processing up to step 4 for all the overlapping groups, if there are non-reassignable pixels, the assignment destination of each non-reassignable pixel (pixel of image k−1) is canceled, Treat as the initial detection target (first detected target) and replace the track score with the value of the pixel likelihood ratio only.

  Next, the end determination unit 8 determines that the integration process is ended when the number of frames subjected to the integration process reaches a predetermined upper limit value, and proceeds to the target detection unit 9. The process returns to the observation image acquisition unit 2. Further, as an end determination condition, it may be determined to end when the elapsed time from the start of integration processing reaches a predetermined upper limit value.

  Next, the target detection unit 9 obtains the average μ and standard deviation σ of the track score from the final result of the integration process, and when the coefficient is Q, the pixel whose track score exceeds the threshold value TH defined by Equation (5) Detect as a target candidate.

  In the target detection apparatus as described above, by suppressing the overlapping allocation of pixels between frames that occur in the integration process and associating with one-to-one correspondence, it is possible to suppress the spread of target pixels that occur by repeating the integration process, There is an effect of improving the target detection accuracy.

Embodiment 2. FIG.
FIG. 5 is a block diagram showing the configuration of the target detection apparatus according to Embodiment 2 of the present invention. The target detection apparatus according to the first embodiment collectively eliminates the overlapping assignment that occurs every time the integration process is executed, whereas the target detection apparatus according to the second embodiment eliminates the overlapping assignment that occurs each time the integration process is performed. Only the pixel with the best track score is finalized, other pixels in the overlapping allocation state are set as the integration target pixels, and only the integration target pixels are newly processed, and the integration processing in the same frame set is not allocated. The difference is that it repeats until there are no more pixels.

  In the block diagram shown in FIG. 5, functions of the sensor unit 1, the observed image acquisition unit 2, the likelihood ratio calculation unit 3, the transition probability calculation unit 4, the duplicate assignment extraction unit 6, the end determination unit 8, and the target detection unit 9. Since the operation is the same as that of the first embodiment, the description is omitted.

  The integration processing unit 5A integrates the likelihood ratio obtained by the likelihood ratio calculation unit 3, the transition probability obtained by the transition probability calculation unit 4, and the track score of the image one frame before, and calculates the image of the current frame. An apparatus that obtains a track score and associates each pixel of the current frame image with a pixel of an image one frame before that maximizes the track score of each pixel of the current frame image. 1 is different from the integration processing unit 5 shown in FIG. 1 except that the integration processing is performed only on the integration processing target set in the integration processing target setting unit 10.

  Based on the processing result of the duplication allocation extraction unit 6, the integration process target setting unit 10 sets the other pixels as the repeated integration process targets in the same frame set except for the pixel with the most track score at each duplicate occurrence location. The process end determination unit 11 determines whether or not to end the repeated execution of the integration process for the same frame set based on the integration process target setting result of the integration process target setting unit 10, and if so, the initial detection processing unit 12. If not, the process returns to the integration processing unit 5A. The initial detection processing unit 12 regards the pixel causing the duplicate assignment as the initial detection pixel when the duplicate assignment is not finally resolved, cancels the assignment, and changes the track score of the pixel.

  Next, portions of the operation of the target detection apparatus in the second embodiment that are different from those in the first embodiment will be described. The operation of the integration processing unit 5A is basically the same as that of the ship draft processing unit 5 shown in the first embodiment, except that the target of the integration processing is limited. The operation of the integration processing unit 5A will be described using an example of processing for a set of an image k-1 and an image k (hereinafter referred to as {k-1, k}).

  The integration processing unit 5A performs integration processing for all pixels (that is, the same as the integration processing unit 5) in the first process for the set of {k-1, k}. Next, in the n-th process (n ≧ 2) for the set of {k−1, k}, the integration process is limited to the integration process target pixels set in the integration process target setting unit 10, and the allocation is confirmed. For the pixels (details are shown in the description of the operation of the integration processing target setting unit 10), the allocation result and the track score obtained so far are maintained.

  Next, the operation of the integration process target setting unit 10 will be described. The integration process target setting unit 10 sets the pixels associated one-to-one between the {k−1, k} pairs as the allocation confirmed pixels and the other as the integration process target pixels. Details are shown below.

step 1. Allocation decision pixel decision (1)
From the processing result of the duplicate allocation extraction unit 6, the allocated pixel of the image k-1 and the pixel of the image k associated with the allocated pixel are determined as allocation confirmed pixels. In FIG. 6, pixel 1, pixel 2, and pixel 8 of image k-1 are assigned pixels, and pixel 1, pixel 2, and pixel 7 of image k are respectively associated with these assigned pixels. , Pixel 1, pixel 2, and pixel 8 of image k-1, and pixel 1, pixel 2, and pixel 7 of image k are assigned fixed pixels. Although the image is originally represented in a two-dimensional array, it is expressed in a one-dimensional array in FIG. 6 for convenience.

step 2. Assignment decision pixel determination (part 2)
From the processing result of the duplicate assignment extraction unit 6, in each duplicate group, the pixel of the image k-1 and the pixel of the image k having the highest track score are selected and set as assignment confirmed pixels.

  In the example of FIG. 6, the pixel 4 of the image k−1 and the pixels 3, 4, and 5 of the image k constitute an overlap group 1, and the pixel 6 of the image k−1 and the pixels 6, 8, and 9 of the image k overlap. Group 2 is configured. In the overlap group 1, if the track score of the pixel 5 is the largest among the pixels of the image k, the pixel 4 of the image k−1 and the pixel 5 of the image k are set as the allocation confirmed pixels. Similarly, in overlap group 2, pixel 6 of image k-1 and pixel 8 of image k are assigned fixed pixels.

step 3. Determining the pixel to be integrated
Pixels other than the allocation-determined pixels determined in step 1 and step 2 are set as integration processing target pixels. In the example of FIG. 6, the pixels 3, 5, 7, and 9 of the image k-1 and the pixels 3, 4, 6, and 9 of the image k are the integration target pixels.

  Next, the operation of the repetition process end determination unit 11 will be described. The repetition process end determination unit 11 calculates the number of pixels to be integrated from the processing result of the integration process target setting unit 10, and when the number of pixels to be integrated is 0, or the number of pixels to be integrated is 1 or more. If the number of pixels subject to integration processing is the same twice, the integration processing is repeated and the process proceeds to the initial detection processing unit 12. In other cases, the process returns to the integration processing unit 5A.

  Next, the operation of the initial detection processing unit 12 will be described. The initial detection processing unit 12 treats the integration target pixels as the initial detection target when the integration target pixels remain after the completion of the {k-1, k} sets of repetition integration processing, and sets the track score to Replace with the value of likelihood ratio only.

  In the target detection apparatus as described above, the processing target is limited to only the pixels where overlap occurs, and the integration process in the same frame set is repeated to eliminate the overlapping allocation of pixels between frames, and the pixels between frames are one-to-one. As a result, the spread of the target pixel is suppressed and the target detection accuracy is improved.

Embodiment 3 FIG.
FIG. 7 is a block diagram showing the configuration of the target detection apparatus according to Embodiment 3 of the present invention. While the target detection apparatus in the first embodiment initially detects the pixels for which the duplicate assignment has not been eliminated, the target detection apparatus in the third embodiment has not finally eliminated the duplicate assignment. The difference is that a connected area of pixels is extracted, one pixel is left in the same connected area, and the track score of another pixel is replaced with the minimum value of the entire image.

  In the block diagram shown in FIG. 7, the sensor unit 1, the observed image acquisition unit 2, the likelihood ratio calculation unit 3, the transition probability calculation unit 4, the integration processing unit 5, the duplicate assignment extraction unit 6, the end determination unit 8, the target Since the function and operation of the detection unit 9 are the same as those in the first embodiment, description thereof is omitted. Also, the function and operation of the duplicate allocation elimination unit 7A are basically the same as those of the duplicate allocation elimination unit 7 shown in the first embodiment, and the difference is that “reassignable pixel processing (initial detection)”. Since only the process is not performed, the description of the operation of the duplicate allocation elimination unit 7A is also omitted.

  The connected region extraction unit 13 extracts a connected region for the overlapping assignment unresolved portion based on the processing result of the overlapping assignment eliminating unit 7A, and the connected region reducing unit 14 extracts the connected region extracted by the connected region extracting unit 13. Replace the track score of all pixels except the pixel with the largest track score in the image with the minimum track score of the entire image.

  Next, the operation of the connected area extracting unit 13 and the connected area reducing unit 14 will be described. First, the connected region extraction unit 13 acquires a location where the duplicate assignment has not been eliminated from the processing result of the duplicate assignment elimination unit 7A. In the example of FIG. 8, the overlap allocation of the pixels 3, 4, and 5 of the image k remains with respect to the pixel 4 of the image k−1 (overlap group 1), and the image k corresponds to the pixel 6 of the image k−1. The overlapping allocation of the pixels 6, 8 and 9 remains (overlapping group 2). Although FIG. 8 is expressed in a one-dimensional array for convenience, an actual image is a two-dimensional array.

  Next, the connected region extraction unit 13 examines connectivity with respect to the pixels of the image k belonging to each overlapping group, and extracts pixels having a connected relationship as connected regions.

  Here, the connection relationship and the definition of the connection area are shown. FIG. 9 shows a connection relationship of pixels, and it is assumed that eight pixels (pixels A to H) around the pixel X are connected to the pixel X. A set of pixels connected by a connection relationship is defined as a connection region. FIG. 10 shows an example of a connected area. Pixels A, B, C, D, and E constitute one connected area (connected area 1), and pixels F, G, and H have different connected areas ( It constitutes a connecting area 2).

  In the example of FIG. 8, the pixels 3, 4, and 5 of the image k in the overlap group 1 constitute a connected region, and the pixels 8 and 9 of the image k in the overlap group 2 constitute a connected region. The pixel 6 of the image k is connected to the pixel 5 of the image k. However, since the overlapping group is different, the pixel 6 of the image k is not included in the connected region.

  Next, the operation of the connected area reduction unit 14 will be described with reference to FIG. The upper part of FIG. 11 shows the processing of the connected region extracting unit 13, and three connected regions of connected regions 1, 2, and 3 are extracted.

  The connected region reduction unit 14 obtains a pixel having a maximum track score in each connected region from the processing result of the connected region extraction unit 13. In the example of FIG. 11, the lower left side shows the extraction result of the maximum track score, and the pixels painted black in each connected region indicate the pixels having the maximum track score.

  Next, the connected area reduction unit 14 replaces the track scores of all the pixels except the pixels having the maximum track score with the minimum track score of the entire image k for each connected area. In the example of FIG. 11, the lower right side shows the track score replacement result. If white, the pixel has a high track score, and if black, the pixel has a low track score. By the processing of the connected area reduction unit 14, only one pixel having a high track score remains in each connected area.

  In the target detection apparatus as described above, overlapping allocation of pixels between frames that occurs in the integration process is canceled and associated with one-to-one, and a connected region is configured for pixels for which overlapping allocation could not be canceled. By leaving only one pixel having the maximum track score from among the pixels and replacing the track score of the other pixels with the minimum value, there is an effect of suppressing the spread of the target pixel and improving the target detection accuracy.

Embodiment 4 FIG.
FIG. 12 is a block diagram showing a configuration of the target detection apparatus according to Embodiment 4 of the present invention. The target detection device in the fourth embodiment is obtained by replacing the initial detection processing unit 12 in the target detection device shown in the second embodiment with a connected region extracting unit 13 and a connected region reducing unit 14, and can perform allocation cancellation. The difference from the target detection apparatus according to the second embodiment is that a connected area reduction process is performed instead of performing the initial detection process on the pixels that have not been detected. Since each device constituting the block diagram shown in FIG. 12 and its operation are the same as those shown in the first embodiment, the second embodiment, and the third embodiment, description of each operation is omitted.

  In the target detection device as described above, the processing target is limited to only the pixels where overlap occurs and the integration process in the same frame set is repeated to eliminate the overlapping allocation of pixels between frames, and further, the overlapping allocation cannot be resolved. For pixels, only one pixel that has the maximum track score is left out of the pixels that make up the connected region, and the track score of other pixels is replaced with the minimum value, thereby suppressing the spread of the target pixel. There is an effect of improving the detection accuracy.

  DESCRIPTION OF SYMBOLS 1 Sensor part, 2 Observation image acquisition part, 3 Likelihood ratio calculation part, 4 Transition probability calculation part, 5, 5A Integration processing part, 6 Duplicate allocation extraction part, 7, 7A Duplicate allocation cancellation part, 8 End determination part, 9 Target detection unit, 10 integration process target setting unit, 11 repetition process end determination unit, 12 initial detection processing unit, 13 connected region extraction unit, and 14 connected region reduction unit.

Claims (4)

A target detection apparatus that enables target detection even in a low SNR environment by integrating a plurality of continuously captured images.
A sensor unit for observing the target;
An observation image acquisition unit that acquires an image observed by the sensor unit;
A likelihood ratio calculation unit for calculating a likelihood ratio of each pixel in the acquired observation image;
A transition probability calculation unit that calculates the transition probability of each pixel in the acquired observation image according to the target assumed motion model,
An integration processing unit that performs an integration process of the likelihood ratio obtained by the likelihood ratio calculation unit, the transition probability obtained by the transition probability calculation unit, and a track score that is a processing result of one frame before;
From the result of the integration processing by the integration processing unit, a plurality of pixels of the current frame image are associated with the same pixel of the image one frame before, and a duplicate allocation extraction unit that extracts a location where duplicate allocation has occurred,
A duplicate assignment eliminating unit for eliminating the duplicate assignment extracted by the duplicate assignment extracting unit;
An end determination unit that determines the end of the integration processing based on the number of integration frames or the integration processing time;
A target detection unit that detects a pixel having a higher track score than the finally obtained processing result, and
The overlapping allocation extracting unit associates a location where overlapping allocation occurs in which a plurality of pixels of the current frame image are allocated to the same pixel of the image one frame before and a pixel of the current frame image. Extract unallocated pixels from the previous image that is not
The duplicate allocation elimination unit leaves a pixel of the current frame image having the maximum track score for each duplicate allocation occurrence location so that the other pixels and the unallocated pixels of the image one frame before correspond one-to-one. If there is a pixel in the current frame image that has been reassigned between pixels and has not been reassigned and the duplicate assignment has not been resolved, the pixel is regarded as the initial detection pixel and the track score is set to the value of the likelihood ratio only. A target detection device characterized by replacing. The target detection apparatus according to claim 1,
In place of the duplicate allocation elimination section,
An integration processing target setting unit for setting an integration processing target for repeatedly executing the integration processing in the same frame set based on the occurrence status of duplicate allocation;
A repetitive process end determination unit for determining the end of the integration process repeatedly executed in the same frame set;
And an initial detection processing unit that regards the pixel of the current frame image as an initial detection pixel when overlapping allocation occurs after the integration processing in the same frame set ends,
The integration processing target setting unit leaves the association between the pixel of the current frame image with the maximum track score and the pixel of the image one frame before at each overlap allocation occurrence location, In addition, a non-assigned pixel that is not associated with each pixel of the current frame image is extracted from the image of the previous frame, and the undetermined pixel and the unassigned pixel are reset as integration processing targets.
The repetition process end determination unit ends the repetition of the integration process within the same frame set when the number of unconfirmed pixels disappears or becomes zero,
The initial detection processing unit, when overlapping allocation portions remain after the integration processing within the same frame set, at each overlapping allocation occurrence location, the remaining portions excluding the pixels of the current frame image having the maximum track score And the track score is replaced with the likelihood ratio of the pixel.
The integration processing unit is configured to perform integration processing on unconfirmed pixels and unassigned pixels when performing repetitive processing within the same frame set. The target detection apparatus according to claim 1,
A connected region extracting unit that extracts adjacent pixels of each overlapping allocation occurrence location as a connected region when the duplicate assignment cannot be completely eliminated by the processing of the duplicate assignment removing unit;
A connected region reduction unit that changes the track score of other pixels except for one of the pixels constituting each extracted connected region;
The connected region reduction unit leaves a pixel having the highest track score for each connected region, and replaces the track score of other pixels with the minimum track score of the entire current frame image. The target detection apparatus according to claim 2,
Instead of the initial detection processing unit,
By repeating the integration process in the same frame set by the repetition process end determination unit, the overlapping allocation of pixels between frames is eliminated, and further, when the overlapping allocation cannot be eliminated, the adjacent pixels where each overlapping allocation occurs are determined. A connected region extraction unit for extracting as a connected region;
A connected region reduction unit that changes the track score of other pixels except for one of the pixels constituting each extracted connected region;
The connected region reduction unit leaves a pixel having the highest track score for each connected region, and replaces the track score of other pixels with the minimum track score of the entire current frame image.

Images