JP6301440B2 - Object detection system and object detection method - Google Patents

Description

  The present invention relates to a detection system and a detection method, and more particularly to an object detection system capable of adaptive learning and a method capable of detecting an object and adaptive learning.

  Advanced Driver Assistance Systems (abbreviated as ADAS) are adopting artificial intelligence to prevent traffic accidents. At present, many companies are focusing on developing technologies such as a lane detection system, a parking assistance system, and a collision avoidance system that apply the machine learning ability of artificial intelligence (see, for example, Patent Document 1).

  In each of the aforementioned systems, the image identification technique is an indispensable technique, and improving the accuracy of image identification is a problem in each of the aforementioned systems. For this reason, an algorithm related to machine learning is mainly constructed, and the parameters of the classifier are mechanically trained thereby, thereby reducing the misjudgment rate of image identification.

  Thus, how to effectively reduce the misjudgment rate in various road environments is an important point in the current research and development of driving assistance systems.

Taiwan Patent No. I478835

  However, since the performance of a vehicle embedded system mounted on a vehicle is limited, the performance of an identification device provided in the vehicle embedded system is also limited.

  Accordingly, an object of the present invention is to provide an object detection system and an object detection method capable of effectively reducing an erroneous determination rate through machine learning even with limited performance of a vehicle embedded system.

  As means for achieving the above object, the present invention provides the following object detection system.

That is, when an image is acquired and whether or not a part of the image is similar to a predetermined object is determined by a classification process using weak classifier parameters, and the image is output and determined to be similar Is a detection unit configured to output an object detection signal;
The detection unit is communicatively connected to receive the image, and a classification process using a weak classifier parameter and a strong classifier parameter are used to determine whether the part of the image is similar to the object. A machine learning unit configured to determine each by a classification process;
The machine learning unit further performs a training process using the image subjected to the determination when the determination based on the weak classifier parameter and the determination based on the strong classifier parameter are different, and performs a new weak classification. An object detection system, wherein the weak classifier parameter of each of the detection unit and the machine learning unit is updated with the new weak classifier parameter. Provide a detection system.

The present invention also provides the following object detection method. That is, an object detection method executed by an object detection system including a detection unit and a machine learning unit,
(A) A step of acquiring an image by the detection unit and performing a classification process using weak classifier parameters to determine whether a part of the acquired image is similar to a predetermined object. When,
(B) outputting a target detection signal by the detection unit when it is determined that a part of the image is similar to a predetermined target;
(C) outputting the image to the machine learning unit by the detection unit;
(D) After the image is received by the machine learning unit, whether or not a part of the received image is similar to an object is classified using a weak classifier parameter, and a strong classifier parameter Performing two determinations by performing each of the classification processes using
(E) When the results of the two determinations are different, the machine learning unit acquires a new weak classifier parameter by performing a training process using the images subjected to the two determinations, And updating a weak classifier parameter of each of the detection unit and the machine learning unit with the new weak classifier parameter.

  According to the above means, by providing the detection unit, classification processing is performed on the acquired image, and it is possible to immediately determine whether the object is included in the image, and further together with the detection unit. By including a machine learning unit, it is mechanically checked whether there is an error in the result of the determination by the detection unit, and if there is an error, training processing is performed on the parameters of the classifier using the image. Since the erroneous determination rate can be reduced by automatically performing the above, the object detection accuracy can be improved efficiently and promptly without requiring manual operations such as auxiliary determination and labeling.

It is a block diagram which shows 1st Embodiment of the target object inspection system which concerns on this invention. It is the schematic which shows embodiment by which the detection unit is arrange | positioned at the vehicle and the machine learning unit is connected to the detection unit so that communication is possible. It is the schematic which shows one Example of the weak classifier and the strong classifier in the target object inspection system which concerns on this invention. It is a block diagram which shows 2nd Embodiment of the target object inspection system which concerns on this invention. FIG. 2 is a schematic diagram showing an embodiment in which both a detection unit and a machine learning unit are arranged in a vehicle. It is a flowchart which shows one Embodiment of the target object detection method which concerns on this invention.

  Hereinafter, specific embodiments of the present invention will be described with reference to the drawings.

<Object detection system>

(First embodiment)

  1 and 2 show a first embodiment of an object detection system according to the present invention. The object detection system includes a detection unit 2 and a machine learning unit 3.

  The detection unit 2 is arranged in a vehicle 9 such as an automobile (see FIG. 2), acquires an image, and determines whether or not a part of the acquired image is similar to a target object using a weak classifier parameter (a set of parameter groups). ), And when it is determined that part of the image is similar to the object, that is, when it is determined that the image includes the object, the object detection signal , And when a new weak classifier parameter (a set of parameter groups) is received, the original weak classifier parameter is updated accordingly. Hereinafter, the structure of the detection unit 2 will be described more specifically.

  In the present embodiment, the image is illustrated as being taken from a moving image that is installed in the vehicle 9 and is recorded by a drive recorder (not shown) that records video around the vehicle 9. 9 is provided to the detection unit 2 for determining the presence or absence of an object in the vicinity of 9. Here, the object is defined in advance in the object detection system, and when the vehicle 9 is an automobile, for example, it is a pedestrian or other vehicle that becomes an obstacle to driving if it is too close. However, the present invention is not limited to this, and the object detection system can be applied to vehicles other than automobiles, and the types of vehicles and objects are not limited to this example.

  The detection unit 2 includes an image acquisition module 21, a first image processing module 22, a first classifier module 23, an output module 24, and a parameter module 25.

  The image acquisition module 21 is configured to receive a moving image shot by the drive recorder and extract an image as a still image from the moving image. In the following description, when an image is simply described, the image acquired by the image acquisition module 21 is indicated.

  The first image processing module 22 is electrically connected to the image acquisition module 21 so as to receive an image, and extracts a first partial image that is a part of the image from the received image. The first partial image is a partial image used for determination by the detection unit 2.

  The first classifier module 23 is electrically connected to the first image processing module 22 so as to receive the first partial image, and performs a classification process using weak classifier parameters (a set of parameter groups). And determine whether the first partial image is similar to the object. When it is determined that the first partial image is similar to the object, the first classifier module 23 determines that the object exists in the vicinity of the vehicle 9 and notifies the driver of the vehicle 9 of the presence of the object. The object detection signal for output is output.

  The output module 24 is electrically connected to the image acquisition module 21 so as to receive an image, and is configured to output the image to the machine learning unit 3.

  The parameter module 25 is electrically connected to the first classifier module 23 and configured to store the weak classifier parameters used by the first classifier module 23 for performing the above-described classification processing. Has been.

  The machine learning unit 3 is communicatively connected to the detection unit 2 so as to receive an image, and performs processing related to classification by using a weak classifier parameter and a strong classifier parameter (both are a set of parameter groups). 2 determinations of whether or not a part of the image is similar to the object (ie, determination by classification processing using weak classifier parameters and determination by classification processing using strong classifier parameters) It is comprised so that each may be performed. If the results of the two determinations are different, training processing is performed using the images, a new weak classifier parameter is obtained, and the parameter module 25 of the detection unit 2 is determined based on the obtained new weak classifier parameter. Update the weak classifier parameters stored in.

  In the present embodiment, the machine learning unit 3 is installed in a remote server (not shown), for example, as shown in FIG. 2, but is not limited thereto, and is installed in the vehicle 9 together with the detection unit 2 as shown in FIG. It may be configured as follows. The machine learning unit 3 is connected to the detection unit 2 wirelessly or by wire so that it can receive images from the output module 24 and transmit new weak classifier parameters to the parameter module 25. Hereinafter, the structure of the machine learning unit 3 will be described more specifically.

  The machine learning unit 3 includes a second image processing module 31, a second classifier module 32, a training module 33, and an update module 34.

  The second image processing module 31 is configured to receive an image from the output module 24 of the detection unit 2 and extract a second partial image including the same contents as the first partial image from the image. . The second partial image is a target of the above-described two determinations of the machine learning unit 3.

  The second classifier module 32 is electrically connected to the second image processing module 31 so as to receive the second partial image and the original image from the detection unit 2, and the weak classification of the machine learning unit 3. The calculation process related to the classification is performed using each of the classifier parameter and the strong classifier parameter, and two determinations are made as to whether or not the second partial image is similar to the object. Here, since the weak classifier parameters used when the second classifier module 32 and the first classifier module 23 execute the above-described classification processing are similar to each other, the weak classification by the second classifier module 32 is performed. The determination result using the classifier parameter is the same as the determination result by the first classifier module 23. In addition, since the total number of strong classifier parameters is larger than the total number of weak classifier parameters, the determination result based on the strong classifier parameters has higher accuracy. Therefore, the determination result based on the strong classifier parameter is regarded as a correct result. If the two determination results performed using the weak classifier parameter and the strong classifier parameter are different from each other, the determination result using the weak classifier parameter is incorrect, and the image in which the erroneous determination has occurred is determined. Output from the second classifier module 32 to the training module 33.

  The training module 33 is electrically connected to the second classifier module 32 so as to receive the image in which the above-described misjudgment occurs, and for training the image of the parameters of the second classifier module 32. As a training sample, a new weak classifier parameter (a set of parameter groups) is obtained (hereinafter also referred to as “training process”) and output to the update module 34.

  The update module 34 is electrically connected to the training module 33 to receive new weak classifier parameters and updates the weak classifier parameters of the second classifier module 32 with the new weak classifier parameters. Above, it is configured to send new weak classifier parameters to the parameter module 25 of the detection unit 2 in order to update the weak classifier parameters used by the first classifier module 23. In the present embodiment, the image in which the above-described erroneous determination has occurred is collected, and the collected image is used to train the parameters of the second classifier module 32, thereby updating the original weak classifier parameters. New weak classifier parameters are obtained.

  In addition to the above-described configuration, the update module 34 calculates a reliability score of a new weak classifier parameter using a predetermined image sample set, and the object (pedestrian) defined in advance in the object detection system. And the other weak classifier parameters may be further configured to determine whether the new weak classifier parameters are actually superior to the original weak classifier parameters. In this case, only when the confidence score of the new weak classifier parameter is greater than the confidence score of the original weak classifier parameter, the update module 34 uses the new weak classifier parameter to update the second classifier module 32. The weak classifier parameters are updated, and new weak classifier parameters are transmitted to the parameter module 25 of the detection unit 2 in order to update the weak classifier parameters of the first classifier module 23.

  In one embodiment, the first classifier module 23 may have a weak classifier 231. The weak classifier 231 is used in combination with weak classifier parameters to perform classification processing. Further, the second classifier module 32 may have a strong classifier 321 as shown in FIG. 3, and the strong classifier 321 is used in combination with the strong classifier parameter to execute the classification process. . In the second classifier module 32, the first to M stages of the strong classifier 321 form a weak classifier 322, and the weak classifier 322 is used in combination with the weak classifier parameters to execute the classification process. The In this way, the second classifier module 32 can output two determination results by the strong classifier 321 and the weak classifier 322, respectively.

  According to this embodiment, the object detection system has the following advantages.

  First, by providing the detection unit 2, a real-time calculation is performed on each image extracted from the moving image output by the drive recorder, and a part of the image is similar to the object, that is, the object exists. If it is determined, an object detection signal for notifying the driver of the presence of the object is output. In addition, by continuously updating the weak classifier parameters, the accuracy of determination by the first classifier module 23 can be further improved.

  Furthermore, in addition to having a machine learning unit 3 communicably connected to the detection unit 2, a second classification for performing a classification process using a weak classifier parameter and a strong classifier parameter in order to obtain two determination results By providing the classifier module 32, it is possible to confirm whether there is an error in the determination result by the first classifier module 23. If it is determined that it is an erroneous determination, the image that caused the erroneous determination is output to the learning module 33. The learning module 33 trains the parameters of the second classifier module 32 based on this image and acquires new weak classifier parameters. The update module 34 then updates the original weak classifier parameters with the new weak classifier parameters. In this way, an image that is a result of erroneous determination by the first classifier module 23 is automatically determined, and this image is used for machine learning, thereby reducing the erroneous determination rate in the next image determination. Since the above procedure is automatically performed in the object detection system according to the present invention, there is no need for auxiliary determination by manual work or tagging, and a large amount of learning samples can be quickly collected and weak classifier parameters. Can be trained. Thus, according to the object detection system according to the present invention, the erroneous determination rate of the detection unit 2 can be quickly reduced, and the accuracy of the object detection is greatly improved, so that the driving safety of the vehicle 9 is improved. Can be made more certain.

  Further, as shown in FIG. 2, the detection unit 2 is provided in a vehicle built-in system (not shown) mounted on the vehicle 9, and a server (not shown) wirelessly connected to the detection unit 2 is provided with a machine. By providing the learning unit 3, the machine learning unit 3 can be provided with the second classifier module 32 having a higher computing capacity without being affected by the performance of the vehicle embedded system. Therefore, by using the strong classifier parameter, it is possible to perform real-time processing related to classification and quickly check whether there is an error in the determination result by the first classifier module 23, and the accuracy of the weak classifier parameter. Can be further improved through training and renewal after the confirmation. However, this is one embodiment, and in another embodiment shown in FIG. 5, for example, the machine learning unit 3 can be arranged in a vehicle built-in system mounted on the vehicle 9 together with the detection unit 2. .

  In addition, the update module 34 is configured to update the original weak classifier parameters only when the latter confidence score is higher with the new weak classifier parameters, so that each weak classifier parameter is updated. As a result, the accuracy of the first classifier module 23 and the second classifier module 32 is further improved. On the other hand, if the erroneously determined image is not suitable for training the parameters of the second classifier module 32, the parameters of the first classifier module 23 and the second classifier module 32 are the new weak classifier parameters. Therefore, the detection performance is prevented from deteriorating.

(Second Embodiment)

  4 and 5 show a second embodiment of the object detection system according to the present invention. The second embodiment is the same as the first embodiment described above, but has the following differences compared to the first embodiment.

  In the present embodiment, the first classifier module 23 of the detection unit 2 further includes a distance calculation module 232. When the weak classifier 231 determines that there is an object in the vicinity of the vehicle 9, the distance calculation module 232 acquires the first partial image described above, and the size of the object included in the first partial image. The distance from the vehicle 9 to the object is calculated based on the above, and the object distance signal indicating the distance is output.

  In the present embodiment, the object detection system is electrically connected to the detection unit 2 so as to receive the object detection signal, and is related to the driving state of the vehicle 9 provided with the detection unit 2. Feedback that is configured to further receive a vehicle signal to be received, a collision signal output when the vehicle 9 collides with an object by a collision sensor (not shown) provided in the vehicle 9, and the object distance signal. A unit 4 is further provided. The feedback unit 4 is configured to output an erroneous determination feedback signal to the output module 24 based on one or more of the vehicle signal, the collision signal, the object distance signal, and the object detection signal. The output module 24 outputs an image to the second image processing module 31 of the machine learning unit 3 in response to receiving the erroneous determination feedback signal.

  In the present embodiment, the vehicle signal indicates both the traveling speed and the amount of movement of the brake pedal. However, the vehicle signal is not limited to this, and may indicate either the traveling speed or the amount of movement of the brake pedal. It is not limited to this embodiment.

  The feedback unit 4 is configured to output an erroneous determination feedback signal in at least one of the following situations.

  The first situation is a situation where the deceleration rate of the traveling speed of the vehicle 9 is equal to or less than a preset deceleration rate threshold while the feedback unit 4 receives the object detection signal, or the feedback unit 4 is the target. While the object detection signal is received, the movement amount of the brake pedal of the vehicle 9 is not more than a preset brake pedal movement amount threshold value.

  In this situation, the vehicle 9 is rapidly decelerated regardless of the determination result by the first classifier module 23 that the object is present (that is, the first partial image is similar to the object). Means that the driver has not depressed the brake pedal. Therefore, in this case, the feedback unit 4 determines that the determination by the first classifier module 23 is an error.

  In the second situation, the feedback unit 4 does not receive the object detection signal, whereas the deceleration rate of the traveling speed of the vehicle 9 is larger than a preset deceleration rate threshold, or the feedback unit 4 is the target. Although the object detection signal is not received, the movement amount of the brake pedal is larger than a preset brake pedal movement amount threshold value.

  In this situation, the vehicle 9 is rapidly detected regardless of the determination result by the first classifier module 23 that the object does not exist (that is, the first partial image is not similar to the object). This means that the vehicle has slowed down or the driver has depressed the brake pedal. Therefore, in this case, as in the first situation, the feedback unit 4 determines that the determination by the first classifier module 23 is an error.

  The third situation is when the feedback unit 4 does not receive the object detection signal but receives a collision signal.

  This situation is due to the fact that the first classifier module 23 did not output the object detection signal despite the presence of the object in the vicinity of the vehicle 9, or the object detection signal from the first classifier module 23. Means that the driver did not have sufficient response time.

  The fourth situation is a situation in which the distance from the vehicle 9 to the object indicated by the object distance signal is less than the braking distance of the vehicle 9 obtained by the calculation described below.

  In this situation, the feedback unit 4 determines that the object is present by the first classifier module 23 and that the distance from the vehicle 9 to the object calculated by the distance calculation unit 232 can be safely braked. It is determined that the vehicle 9 continues to approach the object regardless of whether the braking distance is less than. In this case, the feedback unit 4 determines that the determination by the first classifier module 23 is an error.

  Note that the above-described braking distance is obtained by the following equation.

In the equation, S ′ is a calculated braking distance, v is a preset end speed, which is set to zero here, v _o is an initial speed which is a vehicle speed at the start of braking, and a is a predetermined speed The acceleration obtained based on the braking force of 0.4 g, s is the theoretical braking distance obtained by Equation 1, S is the displacement of the vehicle 9 within the driver's response time, and t is the driver's response time. Yes, practically 0.8 seconds.

  As described above, the second embodiment of the object detection system can obtain the same effects as those of the first embodiment, and further has the following advantages.

  The feedback unit 4 can determine whether or not there is an error in the determination by the first classifier module 23 by receiving the object detection signal, the vehicle signal, the object distance signal, and the collision signal. Further, since the output module 24 is configured to output an image to the machine learning unit 3 only when an erroneous determination feedback signal is received, the image output by the output module 24 is the first classifier module. It becomes an effective training sample related to the misjudgment by 23. In this way, the data throughput required for the second classifier module 32 can be greatly reduced. That is, since the specification requirement for the second classifier module 32 is lowered, the design of the second classifier module 32 can be simplified. Therefore, the second classifier module 32 can be configured by an architecture having a general computing capability. Further, as an example, if the machine learning unit 3 is directly mounted on the vehicle built-in system of the vehicle 9 together with the detection unit 2, not only can the design cost be saved, but also equipment for independently providing the machine learning unit 3 on the vehicle 9 is required. Disappear. Further, if the machine learning unit 3 is directly mounted on the vehicle built-in system of the vehicle 9 together with the detection unit 2, it does not require connection to the remote server as compared with the case where the machine learning unit 3 is provided on the remote server. The parameters of one classifier module 23 can be updated immediately.

  In the object detection system according to the present invention, the detection unit 2, the machine learning unit 3, and the feedback unit 4 may be configured by a hardware method (processor, IC chip, etc.), and are executed by a processor, for example. Note that it may be configured by a software scheme such as code containing commands.

<Object detection method>

  As shown in FIGS. 4 and 6, the object detection method according to the present invention is implemented by the above-described object detection system according to the present invention in one embodiment, and includes the following steps.

  In step S51, the detection unit 2 receives a moving image and extracts an image from the received moving image. Further, the detection unit 2 extracts a first partial image that is a part of the image from the acquired image, and outputs the image to the machine learning unit 3 in response to reception of the erroneous determination feedback signal. In addition, the said moving image was image | photographed with the drive recorder (not shown) which is installed in the vehicle 9 with which the target object detection system is arrange | positioned and records the image | video of the periphery of the vehicle 9, for example.

  In step S52, the detection unit 2 performs processing related to classification using the weak classifier parameters of the detection unit 2, and determines whether or not the first partial image is similar to a predetermined object. . When it is determined that the first partial image is similar to the object, the detection unit 2 outputs an object detection signal.

  In step S <b> 53, the feedback unit 4 receives at least a vehicle signal related to the driving situation of the vehicle 9 and determines whether or not to output an erroneous determination feedback signal to the detection unit 2 based on at least the vehicle signal. Here, the vehicle signal is a signal indicating at least one of the traveling speed of the vehicle 9 and the amount of movement of the brake pedal of the vehicle 9.

  In step S53, the feedback unit 4 may further receive a collision signal output from the collision sensor of the vehicle 9 and an object distance signal indicating the distance from the vehicle 9 to the object. In this case, the feedback unit 5 outputs an erroneous determination feedback signal in at least one of the following situations.

  The first situation is a situation where the feedback unit 4 receives the object detection signal, whereas the deceleration rate of the vehicle speed is equal to or less than a preset deceleration rate threshold, or the feedback unit 4 receives the object detection signal. The situation is that the movement amount of the brake pedal is equal to or less than a preset brake pedal movement amount threshold value although it is received.

  The second situation is a situation where the feedback unit 4 has not received the object detection signal, whereas the deceleration rate of the vehicle speed is greater than a preset deceleration rate threshold value, or the feedback unit 4 receives the object detection signal. This is a situation in which the movement amount of the brake pedal is larger than a preset brake pedal movement amount threshold value although it is not received.

  The third situation is when the feedback unit 4 does not receive the object detection signal but receives a collision signal.

  The fourth situation is a situation in which the distance from the vehicle 9 to the object indicated by the object distance signal is less than the braking distance of the vehicle 9. The braking distance of the vehicle 9 is obtained by the above-described predetermined formula based on the traveling speed.

  When an erroneous determination feedback signal is output to the detection unit 2 by the feedback unit 4, the process proceeds to step S54.

  In step S54, the machine learning unit 3 performs processing related to the classification using the weak classifier parameter and the strong classifier parameter, respectively, and performs two determinations as to whether a part of the image is similar to the object. . When these two determination results are different, training processing is performed using the images, and new weak classifier parameters for updating the weak classifier parameters of the detection unit 2 and the machine learning unit 3 are acquired.

  More specifically, in this step S54, the machine learning unit 3 extracts a second partial image that includes the same contents as the first partial image and is a part of the image from the image, and sets the weak classifier parameter. Then, each of the processes related to the classification is executed using the strong classifier parameters, and two determinations are made as to whether or not the second partial image is similar to the object. If these two determination results are different, the image is used as a sample for training the parameters of the machine learning unit 3, thereby obtaining new weak classifier parameters.

  In step S54, the machine learning unit 3 calculates a confidence score of a new weak classifier parameter. If the calculated reliability score is larger than the reliability score of the original weak classifier parameter, the process proceeds to step S55.

  In step S55, the machine learning unit 3 updates its weak classifier parameter using the new weak classifier parameter, and further uses this new weak classifier parameter for updating the weak classifier parameter of the detection unit 2. Output to the detection unit 2.

  As described above, according to the object detection method of the present invention, adaptive learning is possible, and the same effects as those of the above-described embodiment of the object detection system can be obtained.

  As mentioned above, although preferable embodiment of this invention was described, this invention is not limited to this, A various change is possible in the range which does not deviate from the summary.

  The present invention is applied to various advanced driving support systems in addition to being useful as an object detection system for detecting obstacles during driving by being installed in a car and notifying the driver of the presence of obstacles, for example. It is possible.

2 detection unit 21 image acquisition module 22 first image processing module 23 first classifier module 231 weak classifier 232 distance calculation module 24 output module 25 parameter module 3 machine learning unit 31 second image processing module 32 second Classifier module 321 Strong classifier 322 Weak classifier 33 Training module 34 Update module 4 Feedback unit 9 Vehicle

Claims (18)

An image is acquired, and whether or not a part of the image is similar to a predetermined object is determined by a classification process using weak classifier parameters, and the image is output. A detection unit configured to output an object detection signal;
The detection unit is communicatively connected to receive the image, and a classification process using a weak classifier parameter and a strong classifier parameter are used to determine whether the part of the image is similar to the object. A machine learning unit configured to determine each by a classification process;
The machine learning unit further performs a training process using the image subjected to the determination when the determination based on the weak classifier parameter and the determination based on the strong classifier parameter are different, and performs a new weak classification. The object detection system is configured to acquire a classifier parameter and update the weak classifier parameter of each of the detection unit and the machine learning unit with the new weak classifier parameter. The detection unit is configured to receive a video, and
An image acquisition module for extracting the image from the received video;
A first partial image electrically connected to the image acquisition module to receive the image and configured to extract a first partial image as a part of the image to be determined from the image; 1 image processing module,
Electrically connected to the first image processing module to receive the first partial image, performing the classification process using the weak classifier parameters, wherein the first partial image is the target A first classifier module configured to determine whether the object is similar to an object, and to output the object detection signal when it is determined that the first partial image is similar to the object;
The object of claim 1, further comprising: an output module electrically connected to the image acquisition module to receive the image and configured to output the image to the machine learning unit. Detection system. The machine learning unit is
The image received from the output module of the detection unit, the partial image including the same content as the first partial image from the received image, and the image to be determined by the machine learning unit A second image processing module configured to extract a second partial image as a part of
Classification processing using the weak classifier parameter and classification processing using the strong classifier parameter electrically connected to the second image processing module so as to receive the second partial image and the image A second classifier module configured to output each of the determination target images when the determination results are different and the determination results are different from each other;
New weak classifier parameters that are electrically connected to the second classifier module to receive the image and train the parameters of the second classifier module using the image as a training sample. A training module configured to obtain,
Electrically connected to the training module to receive the new weak classifier parameter, updating the weak classifier parameter of the second classifier module with the new weak classifier parameter; An update module configured to send the new weak classifier parameter to the detection unit to update the weak classifier parameter of the first classifier module of the detection unit with the new weak classifier parameter. The object detection system according to claim 2, further comprising: A feedback unit electrically connected to the detection unit to receive the object detection signal;
The feedback unit is configured to receive a vehicle signal related to a driving state of a vehicle provided with the detection unit and to output an erroneous determination feedback signal to the output module based on the received vehicle signal. And
The object detection system according to claim 3, wherein the output module outputs the image to the second image processing module of the machine learning unit in response to reception of the erroneous determination feedback signal. The vehicle signal is a signal indicating a traveling speed of the vehicle,
The feedback unit is configured so that the feedback unit receives the object detection signal, but the deceleration rate of the traveling speed of the vehicle indicated by the vehicle signal is equal to or less than a preset deceleration rate threshold, and the feedback In a situation where the unit has not received the object detection signal but the deceleration rate of the traveling speed of the vehicle indicated by the vehicle signal is greater than a preset deceleration rate threshold, the erroneous determination The target object detection system according to claim 4 which outputs a feedback signal. The vehicle signal is a signal indicating a movement amount of a brake pedal of the vehicle,
The feedback unit is configured such that the feedback unit receives the object detection signal, but the amount of movement of the brake pedal of the vehicle indicated by the vehicle signal is equal to or less than a preset brake pedal movement amount threshold; In any of the situations where the feedback unit does not receive the object detection signal but the amount of movement of the brake pedal of the vehicle indicated by the vehicle signal is greater than a preset brake pedal movement amount threshold The object detection system according to claim 4, wherein the erroneous determination feedback signal is output. The object is an obstacle to driving the vehicle in proximity to the vehicle,
The feedback unit is configured to further receive at least one of a collision signal output by a collision sensor provided in the vehicle and an object distance signal indicating a distance from the vehicle to the object,
The feedback unit includes a situation in which the feedback unit has not received the object detection signal but has received the collision signal, and a situation in which the distance indicated by the object distance signal is smaller than a braking distance of the vehicle. 5. The object detection system according to claim 4, wherein the erroneous determination feedback signal is output even in any of the situations described above. The vehicle signal is a signal indicating a traveling speed of the vehicle,
The object detection system according to claim 7, wherein the braking distance of the vehicle is calculated based on the traveling speed indicated by the vehicle signal.   The update module calculates the reliability score of the new weak classifier parameter, and when the calculated reliability score is larger than the reliability score of the weak classifier parameter, Updating the weak classifier parameter of the second classifier module with the classifier parameter and the new weak classifier parameter to the detection unit to update the weak classifier parameter of the first classifier module. The object detection system of claim 3, wherein the object detection system is configured to transmit. An object detection method executed by an object detection system comprising a detection unit and a machine learning unit,
(A) A step of acquiring an image by the detection unit and performing a classification process using weak classifier parameters to determine whether a part of the acquired image is similar to a predetermined object. When,
(B) outputting a target detection signal by the detection unit when it is determined that a part of the image is similar to a predetermined target;
(C) outputting the image to the machine learning unit by the detection unit;
(D) After the image is received by the machine learning unit, whether or not a part of the received image is similar to an object is classified using a weak classifier parameter, and a strong classifier parameter Performing two determinations by performing each of the classification processes using
(E) When the results of the two determinations are different, the machine learning unit acquires a new weak classifier parameter by performing a training process using the images subjected to the two determinations, Updating the weak classifier parameters of each of the detection unit and the machine learning unit with the new weak classifier parameters;
The object detection method characterized by including. In the step (A), the detection unit receives a moving image, extracts the image from the received moving image, acquires the image, extracts a first partial image from the acquired image, and extracts the image As part
The object detection method according to claim 10, wherein in the step (B), the object detection signal is output when it is determined that the first partial image is similar to the object. In the step (D), a second partial image including the same content as the first partial image is extracted from the image and made a part of the image,
The object detection according to claim 11, wherein in the step (E), the training process is performed using the image as a training sample, and the new weak classifier parameter is transmitted to the detection unit for the update. Method. The object detection system further comprises a feedback unit,
The steps performed by the feedback unit between the step (B) and the step (C), respectively.
(A) receiving a vehicle signal related to a driving state of a vehicle provided with the detection unit;
(B) outputting a false determination feedback signal to the detection unit based on the received vehicle signal;
The object detection method according to claim 12, wherein in the step (C), the image is output to the machine learning unit in response to reception of the erroneous determination feedback signal. The vehicle signal in the step (a) is a signal indicating a traveling speed of the vehicle,
In the step (b), the feedback unit receives the object detection signal, but the deceleration rate of the traveling speed of the vehicle indicated by the vehicle signal is not more than a preset deceleration rate threshold value; In any of the situations where the feedback unit has not received the object detection signal, but the deceleration rate of the traveling speed of the vehicle indicated by the vehicle signal is greater than a preset deceleration rate threshold value, The object detection method according to claim 13, wherein the erroneous determination feedback signal is output. The vehicle signal in the step (a) is a signal indicating a movement amount of a brake pedal of the vehicle,
In the step (b), the feedback unit receives the object detection signal, but the movement amount of the brake pedal of the vehicle indicated by the vehicle signal is equal to or less than a preset brake pedal movement amount threshold value. Either the situation or the situation where the feedback unit does not receive the object detection signal but the movement amount of the brake pedal of the vehicle indicated by the vehicle signal is larger than a preset brake pedal movement amount threshold value The object detection method according to claim 13, wherein the erroneous determination feedback signal is output in the situation described above. The object is an obstacle to driving the vehicle in proximity to the vehicle,
In the step (a), the feedback unit further outputs at least one of a collision signal output from a collision sensor provided in the vehicle and an object distance signal indicating a distance from the vehicle to the object. Receive
In the step (b), the feedback unit does not receive the object detection signal but receives the collision signal, and the distance indicated by the object distance signal is greater than the braking distance of the vehicle. The object detection method according to claim 13, wherein the erroneous determination feedback signal is output even in any of the small situations. The vehicle signal in the step (a) is a signal indicating a traveling speed of the vehicle,
The object detection method according to claim 16, wherein the braking distance of the vehicle in the step (b) is calculated based on the traveling speed indicated by the vehicle signal.   In the step (E), after the training process is performed by the machine learning unit, the reliability score of the new weak classifier parameter is calculated, and the calculated reliability score is used as the weak classification. Updating the weak classifier parameter of the machine learning unit with the new weak classifier parameter when the confidence score of the classifier parameter is greater than the confidence score, and detecting the new weak classifier parameter for the update in the detection unit The object detection method according to claim 12, wherein the object detection method is transmitted.

Images