JP5241687B2 - Object detection apparatus and object detection program - Google Patents

Description

  The present invention relates to an object detection apparatus and an object detection program for detecting a specific target object from an image.

  For example, in the person tracking method described in Non-Patent Document 1, learning of a person detector is performed online and offline. In this method, after the person detection result is obtained, when performing the next person detection, learning data of positive cases and negative cases is generated from the image region around the detection result, and the learning data is used to generate the preliminary data. A person detector having an optimum parameter is obtained from the person detector database obtained during offline learning (online learning). Then, the next person detection is performed using the person detector having the selected optimum parameter.

Yamashita, Fujikichi, Kaoru, Kawade “Person Tracking by Soft Decision Feature and Online Real Boosting for Posture Changes” Image Recognition and Understanding Symposium (MIRU2008), pp. 12-19, 2008

  Since the conventional person tracking method is configured as described above, it is difficult to perform robust object detection against appearance fluctuations caused by lighting environment and posture fluctuations, and environmental fluctuations under complicated backgrounds. There was a problem of being.

  Specifically, in the case of Non-Patent Document 1, the robustness with respect to appearance variation is optimal for the detection result using the positive case and the negative case of the detection target collected based on the person detection result. A person detector having various parameters is selected from a group of person detectors learned offline in advance. For this reason, it is necessary to perform offline learning for all assumed appearance fluctuations in advance. In addition, when detecting a person in the next frame, a person detector that has been learned online using the current detection result is used. Therefore, if the appearance change of the person between frames is large, the detection fails. .

  In terms of robustness against environmental fluctuations, in general, the complexity of the background is directly related to the difficulty of identifying the person from the background. For this reason, the human detection performance deteriorates under a complicated background that cannot be predicted. Here, in the case of Non-Patent Document 1, since it is impossible to assume under what lighting environment and background the person is at the time of initial detection, the initial person detection necessary for proceeding with online learning is possible. Cannot be performed stably.

  Furthermore, when performing human detection using a person detector that has been learned in advance as in the past, in order to increase the detection rate of persons with appearance fluctuations and environmental fluctuations, for example, whether or not the target person exists in the image This can be dealt with by lowering the threshold value for determination. However, when the threshold value is lowered, there is a problem that the detection rate increases, and at the same time, the false detection rate for detecting an object other than the detection target increases. As described above, generally, when a high detection rate is to be realized with respect to all appearance fluctuations and environmental fluctuations, an increase in the false detection rate is unavoidable.

  The present invention has been made in order to solve the above-described problems, and performs object detection that is robust against appearance fluctuations caused by illumination environment and attitude fluctuations, and environmental fluctuations under complex backgrounds. For the purpose.

An object detection apparatus according to the present invention uses a change area extraction unit that extracts a change area in a video, a non-change area extraction unit that extracts a non-change area in a video, and data of the change area and the non-change area. A learning unit that learns the characteristics of a change area that is considered to contain a detection target object and generates a change target detector, and a likelihood that represents the likelihood of the change area of an arbitrary image area in the video using the change target detector And a likelihood determination unit that detects a detection target object from an arbitrary image region by comparing the likelihood calculated by the target likelihood calculation unit with a threshold value.

The object detection apparatus according to the present invention includes a change area extraction unit that extracts a change area in a video, a non-change area extraction unit that extracts a non-change area in a video, and data of the change area and the non-change area. Using a learning unit that learns the characteristics of a change area that is considered to contain a detection target object and generates a change target detector, and using the change target detector, the likelihood of a detection target object in an arbitrary image area in a video is obtained. A first target likelihood calculating unit that calculates a likelihood to represent, and an object detector that has previously learned the characteristics of the detection target object, and calculates a likelihood that represents the likelihood of the detection target object in the arbitrary image region. The target likelihood calculating unit, the combined likelihood calculating unit that combines the likelihoods calculated by the first and second target likelihood calculating units, and the combined likelihood calculated by the combined likelihood calculating unit are compared with a threshold value. The combined likelihood of detecting the object to be detected from an arbitrary image area It is intended and a tough.

An object detection program according to the present invention includes a computer including a change area extraction unit that extracts a change area in a video, a non-change area extraction unit that extracts a non-change area in a video, and a detection target object. Learning means for learning the characteristics of a change area using data of possible change areas and non-change areas and generating a change target detector, and a change area of an arbitrary image area in a video using the change target detector A target likelihood calculating means for calculating likelihood representing likelihood, and a likelihood determining means for detecting a detection target object from an arbitrary image region by comparing the likelihood calculated by the target likelihood calculating means with a threshold It is.

An object detection program according to the present invention includes a computer including a change area extraction unit that extracts a change area in a video, a non-change area extraction unit that extracts a non-change area in a video, and a detection target object. Learning means for generating a change target detector by learning features of the change area using data of possible change areas and non-change areas, and a detection target of an arbitrary image area in the video using the change target detector Using a first target likelihood calculating unit that calculates a likelihood that represents object-likeness and an object detector that has previously learned the characteristics of the detection target object, the likelihood that represents the detection target object-likeness of an arbitrary image region is calculated. The second target likelihood calculating means, the combined likelihood calculating means for combining the likelihoods calculated by the first and second target likelihood calculating means, and the combined likelihood calculated by the combined likelihood calculating means are set as threshold values. Any image compared to It is intended to function as a joint likelihood determination means for detecting the detection target object from the frequency.

  According to the present invention, the change target detector is assumed to be a detection target object from the fact that the detection target object is often moving in the video, and the change target detector is used using the change region and the non-change region in the video. Since the object is detected using this change target detector, the change area can be extracted robustly to the appearance fluctuation and the environment fluctuation, and the appearance fluctuation caused by the lighting environment and the attitude fluctuation and the complicated It is possible to perform object detection that is robust against environmental fluctuations under a rough background.

  Further, according to the present invention, it is assumed that the detection target object is moving in the video, and the change target is detected using the change region and the non-change region in the video, assuming the change region as the detection target object. Since the object is detected using this change target detector and the pre-learned object detector, the change area can be extracted robustly to appearance fluctuations and environmental fluctuations, and the lighting environment and posture Robust object detection can be performed against appearance fluctuations caused by fluctuations and the like, and environmental fluctuations under a complicated background.

It is a block diagram which shows the structure of the object detection apparatus which concerns on Embodiment 1 of this invention. FIGS. 2A and 2B are diagrams for explaining a change area detection method of the change area detection unit 2 shown in FIG. 1. FIG. 2A shows a case where an optical flow is used, and FIG. 2B shows a case where a background difference is used. It is a figure which shows the example of the change area and non-change area which the change area extraction part 3 and the non-change area extraction part 4 which are shown in FIG. 1 extract from an image. It is a figure explaining the image scanning method using the image pyramid of the image scanning part 8 shown in FIG. It is a block diagram which shows the structure of the object detection apparatus which concerns on Embodiment 2 of this invention. 6A and 6B are diagrams for explaining a change characteristic selection method of the change characteristic selection unit 14 shown in FIG. 5, in which FIG. 6A shows the movement direction characteristic of the change area, and FIG. 6B shows the area / position characteristic of the change area. FIG. 6C shows a case where the aspect ratio characteristic of the change region is used. It is a block diagram which shows the structure of the object detection apparatus which concerns on Embodiment 3 of this invention. In the object detection apparatus according to Embodiment 1 of the present invention, it is a block diagram showing a configuration in which an image scanning unit 8 is replaced with a change target detector 7a.

Embodiment 1 FIG.
1 is a block diagram showing a configuration of an object detection apparatus according to Embodiment 1 of the present invention. The object detection apparatus shown in FIG. 1 includes a video input unit 1, a change region detection unit 2, a change region extraction unit 3, a non-change region extraction unit 4, a learning process that performs a learning process as preprocessing for detecting an object from a video. Data 5, a learning unit 6, a change target detector 7 that performs a detection process for detecting an object in the video after the learning process, an image scanning unit 8, a target likelihood calculation unit (first target likelihood calculation unit) 9, An object detector 10, a target likelihood calculating unit (second target likelihood calculating unit) 11, a combined likelihood calculating unit 12, and a combined likelihood determining unit 13 are provided.

The video input unit 1 receives a video signal output from a camera or the like and outputs it to the change area detection unit 2.
The change area detection unit 2 detects and detects an image area changing from each image (frame) of the video signal using background difference, optical flow, etc. (Hideyuki Tamura “Computer Image Processing” Ohm, 2002). Change area. 2A and 2B are diagrams for explaining a change area detection method of the change area detection unit 2. FIG. 2A shows a case where an optical flow is used, and FIG. 2B shows a case where a background difference is used. As shown in FIG. 2A, the change area detector 2 extracts the optical flow 23 from the current frame 21 (t = t0) and the next frame 22 (t = t1) in the video signal, and the optical flow 23 is extracted. The detection change area 24 is detected by using this. As shown in FIG. 2B, the change area detection unit 2 detects a difference between the background image 25 and the current frame 26 (t = t0) in the video signal as a detection change area 27. Note that the detection change region extraction is robust against appearance fluctuations and environmental fluctuations.

The change area extraction unit 3 extracts the following areas from the detected change areas in the image detected by the change area detection unit 2 as change areas and adds them to the learning data 5. If the reduction ratio of the image pyramid by the image scanning unit 8 to be described later is 1 / s0, and the size of the object detector 10 to be described later is a width w0 pixel and a height h0 pixel, the change region extraction unit 3 includes the change region detection unit 2. From the detection change region of width w pixels and height h pixels obtained in step 1, n is obtained such that h / h0≈w / w0≈s0 ⁿ (n = 0, 1, 2,..., N), and the width w0 × s0 ⁿ pixels, and the height h0 × s ⁿ pixels in the region extracted by the change area.

On the other hand, the non-change region extraction unit 4 extracts a region that does not include any detected change region in the image detected by the change region detection unit 2 as a non-change region and adds it to the learning data 5. Further, the non-change region extraction unit 4 may extract a region that includes the detected change region as a part as a non-change region and add it to the learning data 5. Similarly to the change area extraction unit 3, a non-change area extraction is performed with a reduction ratio of an image pyramid by an image scanning unit 8 described later as 1 / s0 and a size of an object detector 10 described later as a width w0 pixel and a height h0 pixel. The unit 4 changes the size (width w × height h) of the non-change region from the basic size to h = h0 × s0 ⁿ , w = w0 × s0 ⁿ (n = 0, 1, 2,..., N), the position is moved in the image, and a non-change area having an arbitrary position and an arbitrary size is extracted.

  FIG. 3 is a diagram illustrating an example of a change area and a non-change area in an image. In the image illustrated in FIG. 3, the change area extraction unit 3 extracts the change area 31, and the non-change area extraction unit 4 includes non-change areas 41 and 42 that partially include the change area 31, and the change area 31. The non-change areas 43 and 44, which are areas that do not contain any at all, are extracted.

  The learning unit 6 uses the change region (the change region 31 in the example of FIG. 3) as the positive example and the non-change region (the non-change regions 41 to 44 in the example of FIG. 3) as the negative case learning data 5. A change target detector 7 having learned features is generated. The learning unit 6 completes learning when, for example, the number of positive cases accumulated in the learning data 5 becomes a predetermined number or more. As a learning algorithm used by the learning unit 6, for example, Support Vector Machine (Vladimir Vapnik “Statistical Learning Theory” Wiley, NY, 1998) can be used.

  As described above, the change target detector 7 is generated. The reason why the change target detector 7 having learned the characteristics of the change area is generated is that the change area is considered to include the detection target object based on the fact that the detection target object often moves in the video. . Therefore, the change target detector 7 detects a change area including the target object, that is, the target object, from the image. In addition, since the extraction of the change region is robust against appearance fluctuations and environmental fluctuations, the generated change target detector 7 is also robust against appearance fluctuations and environmental fluctuations. Further, the change target detector 7 has a low calculation load. This is because the change target detector 7 plays an auxiliary role of the object detector 10 and hence the required detection accuracy is low, and therefore a processing method with a low calculation load can be used.

Next, an object detection process using the generated change target detector 7 will be described.
The image scanning unit 8 moves the scanning window for each image in the video signal. FIG. 4 is a diagram for explaining an image scanning method of the image scanning unit 8. As shown in FIG. 4, the image scanning unit 8 moves the scanning window (arbitrary image region) 84 with respect to the original image 81 and sequentially performs image reduction processing to generate reduced images 82 and 83. Similarly, the scanning window 84 is moved with respect to the reduced images 82 and 83. In the image reduction process, the reduction ratio of the image pyramid is 1 / s0 as described above.

The target likelihood calculating unit 9 calculates a target likelihood P (v) obtained by applying the change target detector 7 generated by the learning unit 6 to the scanning window that moves the image. Since the change area is assumed to be the target object, the target likelihood P (v) indicates the target object likelihood obtained from the change target detector 7. The method of calculating the likelihood using a detector is the method of Thierry et al. (Thierry Chatau, Vincent Gay-Bellle, Frederickhause, and Jean-Thierry Instimate Int. junction with ECCV, 2006) can be used.
When the target likelihood calculating unit 9 detects the target object from the scanning window, the scanning window moves not only the original image but also the reduced image, so that target objects having different sizes can be detected.
When the target likelihood calculating unit 9 calculates the target likelihood P (v), if the learning unit 6 has not completed learning of the change target detector 7, the target likelihood P (v) is set to “1”. "Is output.

  The object detector 10 is a detector that has been obtained in advance using sample image data including a target object, and is a detector that detects the target object itself. This object detector 10 is an object detector used for conventional object detection, and generally has a high calculation load and is weak against appearance fluctuations and environmental fluctuations. The calculation load of the object detector 10 is high because the total number of detection processes is increased by changing the size of the image as in an image pyramid and performing full scanning within images of different sizes. is there. The object detector 10 calculates a filter output value to which a plurality of filters are applied per detection process, and processes the output value with an identification function. At this time, the filter calculation function is often a complex function with a large amount of calculation. As a result, the object detector 10 has a large number of total detection processes, and the calculation amount at each detection process is also large. This increases the calculation load.

  As a detection algorithm of a target object to be obtained in advance, a technique of Viola et al. , Pp. 734 to 741, October 2003) or the like.

  The target likelihood calculating unit 11 calculates a target likelihood P (D) obtained by applying the object detector 10 obtained in advance to the scanning window moving in the image. This target likelihood P (D) indicates whether or not the target object is included in the image within the scanning window. In the first embodiment, the target likelihood calculation unit 9 applies the change target detector 7 to the same scanning window to calculate the target likelihood P (v), and then the target likelihood calculation unit 11 Finds the target likelihood P (D | v) using the object detector 10 as a conditional probability.

  The combined likelihood calculating unit 12 uses the target likelihood calculated by the target likelihood calculating unit 9 and the target likelihood calculated by the target likelihood calculating unit 11 for the same scanning window, using the combined likelihood P ( D, v) = P (D | v) × p (v) is obtained.

The combined likelihood determining unit 13 outputs the result of the target object “detection” when the combined likelihood calculated by the combined likelihood calculating unit 12 is equal to or greater than a predetermined threshold. On the other hand, when the joint likelihood is equal to or less than the threshold, the result of the target object “not detected” is output.
The change area extraction is robust against appearance fluctuations and environmental fluctuations, and the learning unit 6 generates the change target detector 7 based on the fact that the object to be detected often moves in the video. Therefore, by using this change target detector 7, it is possible to calculate the likelihood of the detection target that is robust to appearance fluctuations and environmental fluctuations focusing on the change area with respect to the scanning window in the image.
Further, after calculating the target likelihood by applying the change target detector 7 to the scanning window, the target likelihood is obtained using the object detector 10 as a conditional probability, and the change target detector 7 and the object detector are obtained. 10 is used to detect the appearance probability and the environment fluctuation without increasing the false detection rate that occurs when only the conventional object detector 10 is used. And robust object detection.

As described above, according to the first embodiment, the change region detection unit 2 that detects the image region that changes from the video input from the video input unit 1 and sets it as the detection change region, and the detection change region of the change region detection unit 2 , A change area extraction unit 3 that extracts a change area in the video, a non-change area extraction unit 4 that extracts a non-change area in the video, and learning data in which the extracted change areas and non-change areas are accumulated 5, the learning unit 6 that learns the characteristics of the change region that is considered to include the detection target object and generates the change target detector 7, and the change target detector 7 is used to display the image scanning unit 8 in the video. The target likelihood calculating unit 9 that calculates the target likelihood that represents the change area from the scanning window that moves the object, and the object detector 10 that has previously learned the characteristics of the detection target object are used to determine the likelihood of the detection target object from the scanning window. Calculate the target likelihood to represent The target likelihood calculating unit 11, the combined likelihood calculating unit 12 that combines the target likelihoods calculated by the target likelihood calculating units 9 and 11, and the combined likelihood calculated by the combined likelihood calculating unit 12 as a threshold value. In comparison, the apparatus is configured to include a joint likelihood determination unit 13 that detects a detection target object from the scanning window.
For this reason, a change target detector 7 that focuses on a change region that is robustly extracted with respect to appearance fluctuations and environmental fluctuations and that may include a detection target, and a pre-learned object detector 10 It is possible to perform object detection using the combined likelihood obtained by combining the target likelihoods obtained from the above, and as a result, the appearance fluctuation caused by the lighting environment and the posture fluctuation, etc., and the environmental fluctuation under a complicated background, etc. On the other hand, robust object detection can be performed.

  The object detector 10 of the first embodiment may be configured to use a partial object detector that detects a partial region of the target object. In the case of this configuration, the target likelihood calculating unit 11 calculates the likelihood obtained by applying the partial object detector to the image within the scanning window of the image scanning unit 8.

Embodiment 2. FIG.
FIG. 5 is a block diagram showing a configuration of an object detection apparatus according to Embodiment 2 of the present invention. The same or corresponding parts as those in FIG. The object detection apparatus according to the second embodiment has a configuration in which a change characteristic selection unit 14 is newly added to the object detection apparatus according to the first embodiment, and is different in that the change characteristic selection unit 14 is provided.

In object detection, there are many cases where the motion of an object to be detected is known in advance. Therefore, the change characteristic selection unit 14 selects only the change region that matches the characteristic of the target object from the change regions detected by the change region detection unit 2 and outputs the selected change region to the change region extraction unit 3.
FIG. 6 is a diagram for explaining a change characteristic selection method of the change characteristic selection unit 14. For example, when a person crossing a pedestrian crossing is detected as a target object, as shown in FIG. 6A, the change characteristic selection unit 14 selects a change area having only a left-right movement with respect to the movement direction characteristic. Choose as.
Further, for example, when detecting a pedestrian walking on the road as a target object from a camera image taken from a distance, as shown in FIG. 6B, the change characteristic selection unit 14 is close to the ground with respect to area / position characteristics. Only the change area detected within the corresponding range is selected as the selection characteristic area 142, and the other change areas are set as the exclusion characteristic areas 143 and 144.
Further, for example, when detecting a pedestrian as a target object, as shown in FIG. 6C, the change characteristic selection unit 14 selects a change area of a predetermined aspect ratio as the selection characteristic area 145 with respect to the aspect ratio characteristic, The other change area is an exclusion characteristic area 146.
Note that the change characteristic selection unit 14 may select a corresponding change area by combining a plurality of such various characteristics.

  The change region selected by the change property selection unit 14 as the selection property region is output to the change region extraction unit 3 and added to the learning data 5 by the change region extraction unit 3. On the other hand, the change area selected as the exclusion characteristic area by the change characteristic selection unit 14 is not output to the change area extraction unit 3 and is not added to the learning data 5. Therefore, the change target detector 7 for detecting the change area including only the detection target can be generated based on the learning data 5 including the change area having the same characteristics as the detection target. By combining this change target detector 7 with the object detector 10, high-precision object detection can be realized.

  As described above, according to the second embodiment, the learning unit 6 includes the change characteristic selection unit 14 that selects a change area having a predetermined characteristic from among the change areas extracted by the change area extraction unit 3. Using the learning data 5 in which the changed region selected by the unit 14 and the non-changed region extracted by the non-changed region extracting unit 4 are accumulated, the features of the changed region that is considered to include the detection target object are learned, and the change target detector 7 was generated. For this reason, the change characteristic selection unit 14 prevents the change area in which an object other than the detection target exists from being added to the learning data 5, so that the change target detector 7 having the characteristics of the change area in which there is a non-target object exists. Can be prevented, and as a result, it is possible to realize the generation of the change target detector 7 with high accuracy. Therefore, it is possible to realize highly accurate object detection that is robust against appearance fluctuations and environmental fluctuations.

Embodiment 3 FIG.
FIG. 7 is a block diagram showing the configuration of the object detection apparatus according to Embodiment 3 of the present invention. The same or corresponding parts as those in FIG. The object detection apparatus according to the third embodiment has a configuration in which a clustering unit 15, a target cluster selection unit 16, and object sample data 17 are newly added to the object detection apparatus according to the first embodiment. Is different.

  The clustering unit 15 performs a clustering process on the change area data output from the change area extraction unit 3 in the learning data 5, and classifies the change areas into a plurality of clusters (= object classes). As a clustering method, for example, k-mean clustering (Hideyuki Tamura “Computer Image Processing” Ohm, 2002) can be used.

  The target cluster selection unit 16 selects a target cluster of the object class to be detected using the object sample data 17 and notifies the learning unit 6 of the change area included in the target cluster in the learning data 5 as a positive example. To do.

  The learning unit 6 uses the non-change region as it is for learning for the negative case in the learning data 5, and uses only the change region notified from the target cluster selection unit 16 for the positive case for learning. 7 is generated.

  In the object sample data 17, sample data of a plurality of object classes are prepared as detection candidates. As described above, by switching the object class classified by the target cluster selection unit 16 according to the object sample data 17, it is possible to correspond to a plurality of object classes. For example, in the second embodiment, the change area corresponding to the person is added to the learning data 5 from the image of FIG. 6C using the aspect ratio characteristic and is added to the learning data 5, and the change area corresponding to the dog is excluded. The characteristic area 146 is excluded from the learning data 5. On the other hand, in the third embodiment, two object classes of human and dog are registered in the object sample data 17, and change areas corresponding to human and dog are distinguished and added to the learning data 5. The learning unit 6 generates a change target detector 7 corresponding to two objects, a person and a dog.

  As described above, according to the third embodiment, the detection is performed using the clustering unit 15 that classifies the change region extracted by the change region extraction unit 3 into the object class by the clustering process and the object sample data 17 indicating the detection target object. A target cluster selection unit 16 that selects a change region classified into an object class corresponding to the target object, and the learning unit 6 extracts the change region selected by the target cluster selection unit 16 and the non-change region extraction unit 4 The change target detector 7 is generated by learning the characteristics of the change region that is considered to include the detection target object using the non-change region data. For this reason, since the target cluster selection unit 16 prevents the change area where the object class other than the detection target exists from being added to the learning data 5, the change target detector having the characteristics of the change area where the other than the target object exists. 7 can be prevented, and as a result, the generation of the change target detector 7 with high accuracy can be realized. Therefore, it is possible to realize highly accurate object detection that is robust against appearance fluctuations and environmental fluctuations.

The object detection apparatus of the first embodiment includes the change target detector 7, the object detector 10, the target likelihood calculation unit 9 that calculates the target likelihood using the change target detector 7, and the object detector 10. A target likelihood calculating unit 11 that calculates a target likelihood using the combined likelihood calculating unit 12 that combines the target likelihoods, and a combined likelihood determining unit 13 that determines a threshold using the combined combined likelihoods. Although it was configured, the object detector 10, the target likelihood calculation unit 11, and the combined likelihood calculation unit 12 are not provided, and the change target detector 7, the target likelihood calculation unit 9, and the target likelihood calculation unit 9 The structure provided with the joint likelihood determination part 13 (likelihood determination part) which determines a threshold value using only object likelihood may be sufficient.
Since there is a possibility that the learning unit 6 learns the change region where the object other than the target object exists as a positive example, in this configuration, the change target detector 7 having the characteristics of the change region where the object other than the target object exists is generated. There is a possibility that. However, since the extraction of the change region itself is robust against environmental fluctuations and appearance fluctuations, object detection using only the change target detector 7 is also robust against environmental fluctuations and appearance fluctuations.

  Note that when this configuration is applied to the second embodiment, the change characteristic selection unit 14 prevents an object other than the detection target from being included in the change area. It is possible to prevent the change target detector 7 having characteristics from being generated, and to generate the change target detector 7 with high accuracy.

  In addition, when this configuration is applied to the third embodiment, the target cluster selection unit 16 prevents objects other than the detection target from being included in the change area. It is possible to prevent the change target detector 7 having characteristics from being generated, and to generate the change target detector 7 with high accuracy.

Moreover, in the object detection apparatus of the said Embodiment 1-3, it can also be set as the structure which replaces the image scanning part 8 with the change object detector 7a. FIG. 8 is a block diagram showing a configuration in which the image scanning unit 8 is replaced with the change target detector 7a in the object detection apparatus according to the first embodiment. The change target detector 7 a has the same function as the change target detector 7. These change target detectors 7 and 7 a are used as auxiliary detectors of the object detector 10. Therefore, the detection accuracy required for the change target detectors 7 and 7a is low, and therefore, processing with a low calculation load can be used.
As described above, the processing of each part of the object detection device is divided into learning processing and detection processing. However, in the case of the object detection device of FIG. It is assumed that the detection process can be applied when 7a is generated. In the detection process, by using the change target detector 7a instead of the image scanning unit 8, it is possible to detect a change region including the target object from the image, that is, the target object with a low calculation load. Then, by combining the change target detection unit 7 with the object detector 10 with higher accuracy, it is possible to realize object detection with a low calculation load and high accuracy.
As described above, when the target object is detected, by combining the change target detection units 7 and 7a, the change target detection unit 7 detects a change region including the target object in the image, that is, the target object with a low calculation load. Can do.
Although illustration is omitted, the same effect can be obtained by replacing the image scanning unit 8 of the object detection devices of the second and third embodiments with the change target detector 7a.

  When the object detection device is configured by a computer, the video input unit 1, the change region detection unit 2, the change region extraction unit 3, the non-change region extraction unit 4, the learning unit 6, the change target detector 7, and the image scanning unit 8 are used. , The object detection program describing the processing contents of the object detector 10, the target likelihood calculating units 9, 11, the combined likelihood calculating unit 12, and the combined likelihood determining unit 13 is stored in the memory of the computer. The CPU may execute an object detection program stored in the memory.

  DESCRIPTION OF SYMBOLS 1 Image | video input part, 2 Change area detection part, 3 Change area extraction part, 4 Non-change area extraction part, 5 Learning data, 6 Learning part, 7, 7a Change object detector, 8 Image scanning part, 9 Object likelihood calculation Part (first target likelihood calculating part), 10 object detector, 11 target likelihood calculating part (second target likelihood calculating part), 12 joint likelihood calculating part, 13 joint likelihood determining part, 14 change Characteristic selection unit, 15 Clustering unit, 16 Target cluster selection unit, 17 Object sample data, 21 Current frame, 22nd frame, 23 Optical flow, 24, 27 Detection change area, 25 Background image, 26 Current frame, 31 Change area, 41-44 non-change area, 141, 142, 145 selection characteristic area, 143, 144, 146 exclusion characteristic area.

Claims (8)

A change area extraction unit for extracting a change area in the video;
A non-change area extraction unit for extracting non-change areas in the video;
A learning unit that learns features of the change region that is considered to include a detection target object using the data of the change region and the non-change region, and generates a change target detector;
A target likelihood calculation unit that calculates a likelihood representing the likelihood of the change region of an arbitrary image region in a video using the change target detector;
An object detection apparatus comprising: a likelihood determination unit that detects the detection target object from the arbitrary image region by comparing the likelihood calculated by the target likelihood calculation unit with a threshold value. A change area extraction unit for extracting a change area in the video;
A non-change area extraction unit for extracting non-change areas in the video;
A learning unit that learns features of the change region that is considered to include a detection target object using the data of the change region and the non-change region, and generates a change target detector;
Using the change target detector, a first target likelihood calculating unit that calculates the likelihood representing the likelihood of the detection target object in an arbitrary image region in the video;
A second target likelihood calculation unit that calculates a likelihood representing the likelihood of the detection target object in the arbitrary image region using an object detector that has previously learned the characteristics of the detection target object;
A combined likelihood calculation unit that combines the likelihoods calculated by the first and second target likelihood calculation units, respectively;
An object detection apparatus comprising: a combination likelihood determination unit configured to detect the detection target object from the arbitrary image region by comparing the combination likelihood calculated by the combination likelihood calculation unit with a threshold value. A change characteristic selection unit that selects a change area having a predetermined characteristic among the change areas extracted by the change area extraction unit,
The learning unit learns the characteristics of the change region using the change region selected by the change characteristic selection unit and the non-change region data extracted by the non-change region extraction unit, and generates a change target detector. The object detection device according to claim 1, wherein the object detection device is a device. A clustering unit that classifies the change region extracted by the change region extraction unit into an object class by clustering processing;
Using a sample data of a detection target object, and a target cluster selection unit that selects a change region classified into an object class corresponding to the detection target object,
The learning unit learns the characteristics of the change region using the change region selected by the target cluster selection unit and the non-change region data extracted by the non-change region extraction unit, and generates a change target detector. The object detection device according to claim 1, wherein the object detection device is a device. A change area extraction means for extracting a change area in a video;
Non-change area extraction means for extracting non-change areas in the video;
Learning means for generating a change target detector by learning features of the change area using data of the change area and the non-change area considered to include a detection target object ;
Using the change target detector, target likelihood calculation means for calculating likelihood representing the change area likelihood of an arbitrary image area in a video,
An object detection program for comparing the likelihood calculated by the target likelihood calculating means with a threshold value and functioning as a likelihood determining means for detecting the detection target object from the arbitrary image region. A change area extraction means for extracting a change area in a video;
Non-change area extraction means for extracting non-change areas in the video;
Learning means for generating a change target detector by learning features of the change area using data of the change area and the non-change area considered to include a detection target object ;
First target likelihood calculating means for calculating a likelihood representing the likelihood of the detection target object in an arbitrary image region in a video using the change target detector;
A second target likelihood calculating means for calculating a likelihood representing the likelihood of the detection target object in the arbitrary image region using an object detector that has previously learned the characteristics of the detection target object;
A combined likelihood calculating means for combining the likelihoods calculated by the first and second target likelihood calculating means,
An object detection program for comparing the combined likelihood calculated by the combined likelihood calculating means with a threshold value to function as a combined likelihood determining means for detecting the detection target object from the arbitrary image region. A change characteristic selection means for selecting a change area having a predetermined characteristic among the change areas extracted by the change area extraction means,
The learning means learns the characteristics of the change area using the change area selected by the change characteristic selection means and the non-change area data extracted by the non-change area extraction means, and generates a change target detector. The object detection program according to claim 5 or 6, wherein Clustering means for classifying the change area extracted by the change area extraction means into an object class by clustering processing;
Using the sample data of the detection target object, the target cluster selection means for selecting the change area classified into the object class corresponding to the detection target object,
The learning means learns the characteristics of the change area using the change area selected by the target cluster selection means and the non-change area data extracted by the non-change area extraction means, and generates a change target detector. The object detection program according to claim 5 or 6, wherein

Images