JP2011257984A - Object detection device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect at an early stage and accurately an object with a high risk for a driver.SOLUTION: An object detection device obtains information such as a picked-up image of surroundings of own vehicle and information which indicates a driving condition and a surrounding condition of own vehicle, and deduces a dangerous area where a pedestrian may exist and a risk of the dangerous area on the basis of the obtained information, a risk for each posture on the basis of the relation between the dangerous area and the posture of a hypothetical pedestrian in the dangerous area, and a comprehensive risk (dangerous area risk x posture risk) by multiplying the risk of the dangerous area and the risk per posture. A priority is given to each of the comprehensive risks (dangerous area risk x posture risk) in the order of decreasing comprehensive risk level. The comprehensive risk (dangerous area risk x posture risk) is selected, and a window image is extracted from a search range, which corresponds to the dangerous area of the comprehensive risk (dangerous area risk x posture risk), in the picked-up image, in order of the priority. An identification model corresponding to the posture of the selected comprehensive risk (dangerous area risk x posture risk) is compared with the window image in order to identify whether or not an object is a pedestrian.

Description

  The present invention relates to an object detection device and a program, and more particularly to an object detection device and a program for detecting an object from a captured image.

  2. Description of the Related Art In recent years, an increasing number of vehicles are equipped with an object detection device that performs image processing on an image around a vehicle imaged by an in-vehicle camera, detects an object such as a pedestrian, and presents a detection result to a driver.

  For example, a plurality of templates having different modes are stored for each distance to an object that can be a discrimination target, and the object to be discriminated from the image is detected by the background difference method or the saliency calculation method, and the object is detected. An object discriminating apparatus has been proposed in which a distance corresponding to the distance is detected, a template corresponding to the distance is selected, and this template is applied to an object detected from within an image as a discrimination target (see Patent Document 1). ).

  For objects with a wide variety of postures such as pedestrians, the accuracy of object detection can be improved by applying templates with different modes as in the object discrimination device described in Patent Document 1. Can do. However, by increasing the number of template types, the accuracy of detection can be further improved, but as the number of templates increases, the processing time increases due to detection by applying all templates. .

  Therefore, based on the determination of the movement direction of the target by the movement direction determination unit, the pedestrian pattern selection unit determines the movement direction determination unit from the pedestrian images for each movement direction stored in advance in the pedestrian pattern storage unit. A pedestrian image with a moving posture in the moving direction is selected, and the recognition processing unit recognizes the pedestrian by comparing the selected pedestrian pattern in the moving direction with the image pattern of the target to be recognized in the camera image. An image recognition apparatus has been proposed (see, for example, Patent Document 2).

  In addition, a plurality of reference patterns (for sky, road, road use, etc.) that are different for each background of the image are stored in the reference pattern database, and the input image using vanishing points and lane recognition by the background region dividing unit. An object candidate region detection device has been proposed in which a reference pattern to be used by a pedestrian candidate region detection unit is selected based on which of the background regions is extracted by the reference pattern selection unit by dividing the background of the pedestrian (for example, And Patent Document 3).

  Further, the image processing unit reads a captured image in units of one frame from the image memory, reads the vehicle speed v, the steering angle α, and the inclination angle β from the storage unit in synchronization with the frame rate, and reads the vehicle speed read by the image processing unit. Based on v, steering angle α, and inclination angle β, referring to the LUT, the parameter for specifying the partial image is acquired, the specified partial image is temporarily stored in the frame memory, and the pedestrian recognition processing unit There has been proposed a vehicle periphery recognition system that reads a partial image stored in a frame memory and determines the degree of similarity between the read partial image and a standard pattern read from a standard pattern portion (see, for example, Patent Document 4).

JP 2007-072665 A JP 2009-237897 A JP 2007-328630 A JP 2006-178852 A

  However, in the image recognition device of Patent Document 2, the pedestrian pattern to be used is selected based on the moving direction of the pedestrian. However, for example, when the pedestrian is standing still trying to cross the road, There is a problem that the applied object cannot be applied because the moving direction cannot be recognized.

  Further, in the object candidate area detection apparatus of Patent Document 3, a model to be used for a road and other areas is selected. However, what is focused on is the diversity of background patterns, and the posture of the object. There is a problem that the robustness to the posture change may be reduced because it is not the diversity of the movement. In addition, since the risk of the area and the posture of the object is not taken into consideration, there is a problem that a pedestrian model that is less dangerous for the driver may be used depending on the search range of the object. .

  Moreover, in the vehicle periphery recognition system of Patent Document 4, the search range of the object is set based on the vehicle speed, the steering angle, etc., but this does not consider the risk such as the possibility of collision. Depending on the range, there is a problem that a pedestrian model that is not dangerous for the driver may be used.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide an object detection apparatus and program capable of detecting an object with a high degree of danger for a driver early and accurately. To do.

  In order to achieve the above object, the object detection apparatus of the present invention includes an acquisition unit that acquires a captured image obtained by imaging the periphery of the host vehicle, information indicating a traveling state of the host vehicle, and information indicating a surrounding state of the host vehicle. Based on the information acquired by the acquisition means, a dangerous area in which the object may exist, a dangerous area estimation means for estimating a dangerous area indicating the degree of danger in the dangerous area, and the dangerous area Based on the relationship between the positional relationship with the host vehicle and the posture of the object assumed to be in the dangerous area, the posture-specific risk estimation means for estimating the posture-specific risk indicating the degree of danger for each posture. And, based on the area risk level and the posture-specific risk level, estimate a total risk level for each posture of the object that is assumed to exist in the risk area for each risk area, and Select the danger areas in descending order. And extracting means for extracting a window image from the area on the captured image corresponding to the selected dangerous area, the window image extracted by the extracting means, and the posture of the object in order to identify the object in advance. Identification for identifying whether or not the window image is an image representing an object based on an identification model selected according to the posture of the object assumed to exist in the selected dangerous area from the generated identification model And means.

  According to the object detection device of the present invention, the acquisition unit acquires a captured image obtained by imaging the periphery of the host vehicle, information indicating the traveling state of the host vehicle, and information indicating a surrounding state of the host vehicle, and a dangerous region estimation unit. However, based on the information acquired by the acquisition means, the risk area in which the object may exist and the area risk level indicating the degree of danger in the risk area are estimated, and the posture-specific risk level estimation means Then, based on the relationship between the positional relationship between the dangerous area and the host vehicle and the posture of the object assumed to be in the dangerous area, the risk by posture indicating the degree of danger for each posture is estimated. Then, the extraction means estimates the overall risk level for each posture of the object assumed to exist in the risk area for each risk region based on the area risk level and the posture-specific risk level. A risk area is selected in descending order of the degree of risk, a window image is extracted from an area on the captured image corresponding to the selected risk area, and an identification means identifies a target object in advance from the window image extracted by the extraction means. Whether the window image represents an object based on the identification model selected according to the posture of the target object that is assumed to be in the danger area selected from the identification model generated for each posture of the target object Identify whether or not.

  Thus, in order to set the search range in descending order of the overall risk based on the risk for each region and the risk for each posture of the object, to select an appropriate identification model from the posture-specific identification model, An object with a high degree of danger for the driver can be detected early and accurately.

  Further, the dangerous area estimation means may exclude a dangerous area whose area risk is smaller than a predetermined area risk threshold from a processing target. Further, the posture-specific risk estimation means can exclude postures whose posture-specific risks are smaller than a predetermined posture-specific risk threshold from the processing target. Further, the extracting means excludes the posture of the target object that is assumed to exist in the risk area for each risk area whose total risk level is smaller than a predetermined risk level threshold. Can do. Further, when it is assumed that there is an object in the dangerous area, the dangerous area estimating means can exclude an area where a collision between the object and the host vehicle is inevitable from the dangerous area. Thereby, even when there is a limit to the computational resources, it is possible to preferentially detect an object with a high degree of risk to be detected.

  Further, the dangerous area estimation means can increase the area risk of the dangerous area corresponding to the window image that has been identified as an image representing the object by the identification means in the past. Further, the posture-specific risk estimation means includes the posture of the target object that is assumed to exist in the risk area corresponding to the window image that has been identified as an image representing the target object by the identification unit in the past. It is possible to increase the risk by posture of the posture of the object indicated by the window image. By feeding back the identification result in this way, the estimation accuracy of the region risk and the posture-specific risk is improved.

  In addition, when the object is a pedestrian, the posture of the object includes the orientation of the pedestrian, the degree of open legs, the relationship between hands and feet, the state of standing still, walking or running, and height It can comprise so that it may contain.

  The information indicating the traveling state of the host vehicle can be configured to include a speed, a steering angle, and a posture angle of the host vehicle.

  The information indicating the surrounding situation of the own vehicle includes the position of an object existing around the own vehicle, information indicating whether the object is moving or stationary, map information, and weather around the own vehicle. It can be configured to contain information.

  In addition, the object detection program of the present invention uses a computer to acquire an image obtained by capturing an image of the periphery of the host vehicle, information indicating a traveling state of the host vehicle, and information indicating a surrounding state of the host vehicle, the acquiring unit. Based on the information acquired by the above, a dangerous area in which the object may exist, a dangerous area estimating means for estimating a dangerous area indicating the degree of danger in the dangerous area, and the position of the dangerous area and the host vehicle A posture-specific risk estimation means for estimating a posture-specific risk indicating the degree of danger for each posture based on the relationship between the relationship and the posture of the object assumed to exist in the risk region; Then, based on the risk by posture, the overall risk for each posture of the object assumed to exist in the risk area is estimated for each risk area, and the risk is calculated in descending order of the total risk. Select an area Extraction means for extracting a window image from the area on the captured image corresponding to the selected danger area, and the window image extracted by the extraction means, and generated for each posture of the object in advance to identify the object As an identification means for identifying whether the window image is an image representing an object based on an identification model selected according to the posture of the object assumed to exist in the selected dangerous area from the identified identification model It is a program to make it function.

  The storage medium for storing the program of the present invention is not particularly limited, and may be a hard disk or a ROM. Further, it may be a CD-ROM, a DVD disk, a magneto-optical disk or an IC card. Furthermore, the program may be downloaded from a server or the like connected to the network.

  As described above, according to the present invention, the search range is set in descending order of the overall risk level based on the risk level for each region and the risk level for each posture of the object. Since an easy identification model is selected, it is possible to quickly and accurately detect an object with a high degree of danger for the driver.

It is a block diagram which shows the structure of the target object detection apparatus which concerns on this Embodiment. It is a figure which shows an example of a dangerous area map. (A) It is a figure for demonstrating the case where the risk according to posture is low, and (B) the case where the risk according to posture is high. It is a flowchart which shows the content of the target object detection processing routine in the target object detection apparatus of this Embodiment. It is a figure which shows an example of the estimation result of area | region risk, the risk according to attitude | position, and a comprehensive risk.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, a case will be described in which the present invention is applied to an object detection device that is mounted on a vehicle and detects a pedestrian as an object.

  As shown in FIG. 1, the object detection device 10 according to the present embodiment includes an imaging device 12 that captures a range including an identification target region, a traveling state detection unit 14 that detects the traveling state of the host vehicle, and the host vehicle. A surrounding state detection unit 16 that detects a surrounding state of the computer, a computer 18 that executes an object detection processing routine for detecting a pedestrian based on a captured image output from the imaging device 12, and a processing result in the computer 18 And a display device 20 for displaying.

  The imaging device 12 captures an area including an identification target region and generates an image signal (not shown), and A / D conversion that converts an analog image signal generated by the imaging unit into a digital signal And an image memory (not shown) for temporarily storing the A / D converted image signal.

  The traveling state detection unit 14 includes a vehicle speed sensor that detects the speed of the host vehicle, a steering angle sensor that detects a steering angle, and a gyro sensor that detects a posture angle (pitch angle, yaw angle, roll angle) of the host vehicle. It is configured. Detection values detected by the sensors are input to the computer 18.

  The surrounding state detection unit 16 irradiates the front of the host vehicle while scanning with a laser, and detects a two-dimensional position of an object irradiated with the laser by reflection of the laser, based on satellite signals from GPS satellites. A GPS database that detects the position of the host vehicle, a map database that electronically holds map information including road shapes (eg, linearity, curvature, intersection, etc.) and road types (eg, urban areas, expressways, etc.), and It includes a weather sensor that includes a clear rain sensor, a temperature sensor, a humidity sensor, and the like. Detection values detected by each device and each sensor are input to the computer 18.

  The computer 18 is a CPU that controls the entire object detection apparatus 10, a ROM as a storage medium that stores a program for an object detection processing routine, which will be described later, a RAM that temporarily stores data as a work area, and a connection between them. It is configured to include a bus. In the case of such a configuration, a program for realizing the function of each component is stored in a storage medium such as a ROM or HDD, and each function is realized by executing the program by the CPU. To.

  When the computer 18 is described with function blocks divided for each function realizing means determined based on hardware and software, as shown in FIG. 1, the captured image, the running state of the own vehicle, and the surrounding situation of the own vehicle are shown. A risk region estimation unit 22 that estimates the risk region and the region risk level of the risk region, and a posture-specific risk estimation unit 24 that estimates a pedestrian risk level that is assumed to be in the risk region; A search range setting unit 26 that sets a search range for searching for a pedestrian on the captured image, and sets a priority based on the overall risk based on the area risk and the posture-specific risk, and images in order of priority. A window image extraction unit 28 for extracting a window image of a predetermined size from an area where an image search range is set, and an identification model storage unit 30 in which an identification model for each pedestrian's posture is stored. , By comparing the window image extracting unit by the window image and the orientation was extracted by 28 identification model, a pedestrian identification unit 32 identifies a pedestrian from the captured image can be expressed by the inclusive configure.

  The dangerous area estimation unit 22 acquires the detection values of the traveling state detection unit 14 and the surrounding situation detection unit 16, and detects the position of an object existing around the host vehicle as shown in FIG. In addition, together with the recognition results obtained by image recognition processing of captured images, the types of those objects (preceding vehicles, oncoming vehicles, buildings, stopped vehicles, guardrails, curbs, street trees, utility poles, lanes, pedestrian crossings, etc., and each object Whether the moving object is moving or stationary. Based on these pieces of information, a dangerous area where a pedestrian may exist is estimated. Also, for each dangerous area, estimate the pedestrian existence probability in that area, and the area risk indicating the possibility of collision with the host vehicle after a few seconds assuming that there is a pedestrian in that dangerous area. . Specifically, the distance between the own vehicle and the dangerous area, the position of the dangerous area, the speed of the own vehicle, the steering angle, the road shape, etc. The area risk is estimated for each dangerous area based on information such as whether the object in the vicinity of the dangerous area or the object forming the dangerous area is a moving object or a stationary object, an urban area, or an expressway.

  For example, as shown in FIG. 2, when the areas A to F are estimated as the dangerous areas, the area in front of the own vehicle is more likely to collide when an object is present in the area as it is closer to the own vehicle. Since the possibility increases, the area risk is increased. In particular, area A is in the lane in the traveling direction of the host vehicle, and increases the area risk. Area B is a sidewalk close to the host vehicle, but there is a guardrail between the host vehicle and area B, so even if there are pedestrians in area B, the risk of collision is low. Reduce the area risk. However, as in area C, since there is a high possibility that a pedestrian crosses the gap at the guardrail break, the area risk is increased. Since the area F is around the stopped vehicle, the probability of pedestrians jumping out from the blind spot of the stopped vehicle and getting on and off the stopped vehicle is increased, so the area risk is increased. In addition, since the probability of existence of a pedestrian is naturally high in a region near the position where a pedestrian is detected in the detection process in the previous frame, the region risk is increased. For example, when a pedestrian facing left is detected in the previous frame within the broken line in FIG. 2, the pedestrian existence probability in the area E increases in the current frame, and thus the area risk of the area E is increased.

  Further, when the area risk is smaller than a predetermined area risk threshold, the risk area estimation unit 22 excludes the area from the risk area. In addition, when a pedestrian is present, an area where a collision with the host vehicle is inevitable is excluded from the dangerous area. As in area X in FIG. 2, when the distance from the host vehicle is extremely close and a pedestrian is present in this area, there is a region where a collision with the host vehicle is inevitable. If a pedestrian is detected by the past detection result of this device and the driver is warned, etc., a route is taken so that the pedestrian does not enter this inevitable area by, for example, avoiding the pedestrian already. The possibility that a pedestrian suddenly appears in an inevitable area is extremely low. That is, when there is a pedestrian in the inevitable area, the danger is quite high, but since this inevitable area has a very low probability of pedestrians, it can be excluded from the pedestrian search range.

  The posture-specific risk estimation unit 24 estimates the posture-specific risk in the dangerous region based on the relationship between the positional relationship between the dangerous region and the host vehicle and the posture of the pedestrian assumed to be in the dangerous region. . For example, as shown in FIG. 3A, when the posture of the pedestrian existing on the left sidewalk of the host vehicle is backward, this indicates that the pedestrian is traveling in parallel to the road. The risk of collision with the vehicle is low. On the other hand, even if a pedestrian is also present on the left sidewalk of the own vehicle, as shown in FIG. The possibility of crossing is high, and the risk of collision with the vehicle increases. In this way, different posture-specific risks are determined depending on the relationship between the positional relationship between the host vehicle and the dangerous area and the posture of the pedestrian. In this embodiment, the posture of the pedestrian is directed leftward facing the left hand in the traveling direction when viewed from the driver, rightward facing the right hand in the traveling direction viewed from the driver, and directed in the traveling direction when viewed from the driver. Forward and backward facing the direction of travel.

  For example, since the area A in FIG. 2 is in the road ahead of the host vehicle, it is dangerous for pedestrians of any posture, but this area, which is a roadway, has a backward or forward posture. Pedestrians are unlikely to exist, and there is a high probability of pedestrians crossing this area, that is, pedestrians facing right or left, and if such pedestrians exist, Since the possibility of a collision is high, the risk for each right or left posture is increased. Since the area C is an area on the left side of the host vehicle, when there is a pedestrian who is going to cross to the right, the risk increases, so the risk by posture in the right direction is increased. On the contrary, since the area D is an area on the right side of the host vehicle, the degree of risk for each posture in the left direction is increased. In addition, when a pedestrian is detected in the detection process in the previous frame, there is a high possibility that there is a pedestrian in the posture detected in the previous frame. Increase the risk by posture. For example, if a left-handed pedestrian is detected in the previous frame within the broken line in FIG. 2, the probability that a left-handed pedestrian exists in the area E increases in the current frame. Make it high.

  In addition, the posture-specific risk estimation unit 24 excludes postures whose posture-specific risks are smaller than a predetermined posture-specific risk threshold from the processing target.

  The search range setting unit 26 multiplies the region risk estimated by the dangerous region estimation unit 22 by the posture-specific risk estimated by the posture-specific risk estimation unit 24 to obtain a comprehensive information for each risk region. The risk level is estimated. For example, when the area risk level and the posture-specific risk level are expressed in 5 levels of 1 to 5, respectively, and the risk level increases as the number increases, the area risk level of the area C in the example of FIG. Is 4 and the risk by posture in the right direction in area C is 5, the overall risk is 20. Priorities are set in descending order of the overall risk level, and an area on the captured image corresponding to the risk area is set as a search range. Even in the same danger area, the overall risk may differ depending on the posture of the pedestrian, so the area set once as the search range may be set as the search range again for a different attitude. .

  In addition, the search range setting unit 26 excludes a case where the overall risk is smaller than a predetermined risk threshold from the processing target.

  The window image extraction unit 28 sets a predetermined size of a window (referred to as a search window) from a region set as the search range of the captured image according to the priority for each step. The image is cut out while moving only by the designation. Here, the cut image is referred to as a window image.

  The identification model storage unit 30 stores an identification model that is generated by learning in advance and is referred to when the pedestrian identification unit 32 identifies a pedestrian. The identification model is generated and stored for each posture of the pedestrian (here, leftward, rightward, forward, backward).

  The pedestrian identification unit 32 reads an identification model corresponding to the posture of the pedestrian in the search range selected according to the priority from the identification model storage unit 30, and compares the extracted window image with the window image to identify the pedestrian. It is identified whether or not it is an image to be shown. As the identification method, a known method such as template matching or SVM (support vector machine) can be used. In addition, the pedestrian identification unit 32 controls the display device 20 to display the identification result superimposed on the captured image, and temporarily stores the identification result in a predetermined storage area for use in the processing of the next frame. The stored identification result is fed back to the dangerous area estimation unit 22 and the posture-specific risk estimation unit 24.

  Next, an object detection processing routine executed by the computer 18 of the object detection device 10 of the present embodiment will be described with reference to FIG.

  In step 100, the captured image captured by the imaging device 12, each detection value detected by the traveling state detection unit 14 and the surrounding state detection unit 16 is acquired. Also, the previous frame identification result stored in the predetermined area is acquired.

  Next, in step 102, based on the captured image acquired in step 100, each detected value, and the identification result of the previous frame, the position of an object existing around the host vehicle is detected, and the captured image is subjected to image recognition processing. The types of these objects are discriminated together with the recognized results. Based on these pieces of information, a dangerous area where a pedestrian may exist is estimated, and a dangerous area map as shown in FIG. 2 is generated, for example.

  Next, in step 104, the position of the host vehicle and the dangerous region after a predetermined time predicted based on the distance between the host vehicle and the dangerous region, the position of the dangerous region, the speed of the host vehicle, the steering angle, the road shape, etc. The area risk is estimated for each risk area based on information such as whether the object in the vicinity of the danger area or the object forming the danger area is a moving object or a stationary object, an urban area, or an expressway. For example, in the case where the dangerous area map as shown in FIG. 2 is generated, if the area risk level is expressed in five levels 1 to 5 and the risk level increases as the number increases, FIG. As shown, the area risk “5” in area A, the area risk “2” in area B, the area risk “4” in area C,... Can be estimated.

  Next, in step 106, when the region risk estimated in step 104 is smaller than a predetermined region risk threshold, the region is excluded from the risk region. For example, in the example of FIG. 5, when the area risk threshold is “3”, the area B with the area risk “2” is excluded from the risk area. In addition, when there is a pedestrian, an inevitable area where a collision with the host vehicle is inevitable is excluded from the dangerous area.

  Next, in step 108, based on the relationship between the positional relationship between the dangerous area and the host vehicle and the posture of the pedestrian assumed to be in the dangerous area, the risk by posture in the dangerous area is estimated. For example, in the example of FIG. 2, when the risk by posture is expressed in five levels of 1 to 5 and the risk increases as the number increases, as shown in FIG. It is possible to estimate such as a left posture-specific risk “5”, a right posture-specific risk “5”, a forward posture-specific risk “1”, and a backward posture-specific risk “1”. As for other dangerous areas, the risk by posture is estimated in relation to the dangerous area. Since area B and area X have already been excluded from the dangerous area, the processing of this step for area B and area X is not performed.

  Next, in step 110, postures whose posture-specific risks estimated in step 108 are smaller than a predetermined posture-specific risk threshold are excluded from the processing targets. For example, in the example of FIG. 5, when the posture-specific risk threshold is “2”, the forward and backward directions in the area A with the posture-specific risk “1” are excluded from the processing target.

  Next, in step 112, the region risk estimated in step 104 and the posture-specific risk estimated in step 108 are multiplied to estimate a total risk for each posture for each risk region. . Note that the processing in this step is not performed for areas that have already been excluded from the dangerous area and postures that have been excluded from the processing target. Then, an area on the captured image corresponding to the dangerous area is searched so that the priority becomes higher in the order of higher overall risk, excluding the case where the overall risk is smaller than a predetermined risk threshold. Set as a range. Hereinafter, each posture in the dangerous area is represented as “dangerous area × posture (for example, area A × leftward)”.

  In the example of FIG. 5, for example, the total risk level of area A × leftward is the area risk level “5” of area A × left-side posture-specific risk level “5” of area A = 25. When the risk threshold is “8”, the area F × forward with the overall risk “6” is excluded from the processing target. Priorities i (i = 1, 2,..., N: N is the risk region to be processed × total number of postures) are assigned to the other dangerous regions × postures in descending order of overall risk. If the overall risk is the same, set a priority such as increasing the priority of the person with higher area risk or increasing the priority of the person with higher risk by posture. decide. Here, area A × priority 1 left, area A × right 2 priority, area C × right 3 priority, area F × left 4 priority, area C × forward 5 priority,...・ Give priority as follows.

  Next, in step 114, 1 is set to the variable i representing the priority i. Next, in step 116, the danger area x posture of the priority i is selected. Then, a window image is extracted from the search range on the captured image corresponding to the selected risk area of priority i × posture risk area. The search range on the captured image corresponding to the dangerous area is based on the estimated size and position of the dangerous area, the range corresponding to the dangerous area on the captured image, or the range in which the margin is provided. can do. A window image is extracted while scanning the search range with the search window.

  Next, in step 118, the identification model corresponding to the selected risk area of the priority i × posture is read from the identification model storage unit 30 and compared with the window image extracted in step 116. Is an image showing a pedestrian. For example, for area A × leftward, a leftward identification model is used.

  Next, in step 120, it is determined whether or not the window image is identified as an image showing a pedestrian as a result of the identification in step 118. If it is a pedestrian, the process proceeds to step 122. If it is not a pedestrian, step 122 is skipped and the process proceeds to step 124.

  In step 122, the position and size of the window image and the orientation of the used identification model are stored in a predetermined storage area as an identification result. This identification result is acquired as the identification result of the previous frame in step 100 in the processing of the next frame. Further, based on the identification result, the display device 20 is controlled so that the detected pedestrian is displayed surrounded by a window with respect to the captured image. At this time, an indication such as an arrow indicating the direction of the pedestrian may be performed together.

  In step 124, it is determined whether or not the variable i has been set to N, and it is determined whether or not all the dangerous areas × postures have been set as search ranges. If i ≠ N, the process proceeds to step 126, the variable i is incremented by 1, the process returns to step 116, and the process is repeated. If i = N, the process ends.

  As described above, according to the object detection device of the present embodiment, the search range is set in descending order of the overall risk based on the region risk and the posture risk, and the posture-specific identification model is used. Since an appropriate identification model is selected, a pedestrian with a high degree of danger for the driver can be detected early and accurately.

  In addition, even if the area risk, the posture risk, and the overall risk are smaller than the thresholds set for each, and the inevitable area is excluded from the processing target, the calculation cost is reduced even when there are limited calculation resources. This makes it possible to detect pedestrians with high priority to be detected early and accurately.

  In the present embodiment, a case has been described in which the object to be detected is a pedestrian, but the object may be a two-wheeled vehicle or the like. In the case of a two-wheeled vehicle, the posture can be the orientation, whether it is stationary or traveling, and so on.

  Further, in the present embodiment, the case where each risk level is smaller than the threshold value has been described as being excluded from the processing target. However, when there is a margin in the calculation resources, all the risk areas × postures estimated without being excluded You may make it process about. Further, only the exclusion based on the area risk, only the exclusion based on the posture-specific risk, or only the exclusion based on the overall risk may be performed.

  Further, in the present embodiment, the case where the posture of the pedestrian is leftward, rightward, forward, and backward has been described. However, a more detailed direction may be applied, the degree of open legs, the hands and feet. It may be subdivided according to whether it is stationary, walking or running, adult or child (height). Since the magnitude of pedestrian movement can be ascertained from the degree of leg opening and the relationship between hands and feet, the degree of risk for each posture may be increased when the degree of leg opening is large. Also, when walking rather than standing, it is better to increase the risk by posture when running than when walking. In addition, since children often have less attention to vehicles than adults, it is better to increase the risk by posture in the case of children (when the height is low).

  Further, in the present embodiment, the case has been described in which the overall risk level is estimated by multiplying the area risk level and the posture risk level. However, the overall risk level is determined based on the area risk level and the posture risk level. May be estimated as the sum or average of the above, or each may be weighted to take the sum, product, average, or the like.

DESCRIPTION OF SYMBOLS 10 Object detection apparatus 12 Imaging device 14 Running state detection part 16 Peripheral condition detection part 18 Computer 20 Display apparatus 22 Risk area estimation part 24 Attitude risk degree estimation part 26 Search range setting part 28 Window image extraction part 30 Identification model memory | storage part 32 Pedestrian identification part

Claims (12)

An acquisition means for acquiring a captured image obtained by imaging the periphery of the host vehicle, information indicating a traveling state of the host vehicle, and information indicating a surrounding state of the host vehicle;
Based on the information acquired by the acquisition means, a dangerous area in which there is a possibility that the object exists, and a dangerous area estimation means for estimating an area risk indicating the degree of danger in the dangerous area;
Based on the positional relationship between the dangerous area and the host vehicle and the posture of the object assumed to be in the dangerous area, the posture-specific risk level indicating the degree of danger for each posture is estimated. Risk estimation means;
Based on the risk level of the area and the risk level for each posture, a total risk level is estimated for each posture of the target object that is assumed to be present in the risk area, and Extracting means for selecting the dangerous area in descending order and extracting a window image from an area on the captured image corresponding to the selected dangerous area;
Select according to the posture of the object assumed to exist in the selected dangerous area from the window image extracted by the extraction means and the identification model generated for each posture of the object in advance to identify the object Identification means for identifying whether the window image is an image representing an object based on the identified model;
An object detection apparatus including:   The target object detection apparatus according to claim 1, wherein the dangerous area estimation unit excludes a dangerous area whose area risk is smaller than a predetermined area risk threshold from a processing target.   The target object detection apparatus according to claim 1, wherein the posture-specific risk estimation unit excludes postures whose posture-specific risks are smaller than a predetermined posture-specific risk threshold from processing targets.   The said extraction means excludes the attitude | position of the target object assumed that it exists in this danger area for every said danger area whose said total danger level is smaller than the predetermined risk threshold value from a process target. 4. The object detection device according to any one of items 3.   The dangerous region estimation means, when assuming that an object is present in the dangerous region, excludes from the dangerous region a region where a collision between the object and the host vehicle is unavoidable. The target object detection apparatus of Claim 1.   The risk region estimation unit increases the region risk level of the risk region corresponding to the window image that has been identified as an image representing the object by the identification unit in the past. The target object detection apparatus of description.   The posture-specific risk estimation means includes the window among the postures of the object assumed to exist in the danger area corresponding to the window image that has been identified as an image representing the object by the identification means in the past. The target object detection apparatus according to any one of claims 1 to 6, wherein the risk level according to the posture of the target object indicated by the image is increased.   When the object is a pedestrian, the posture of the object includes the orientation of the pedestrian, the degree of open legs, the relationship between hands and feet, the state of standing still, walking or running, and the height. The target object detection apparatus of any one of Claims 1-7.   The object detection device according to any one of claims 1 to 8, wherein the information indicating the traveling state of the host vehicle includes a speed, a steering angle, and a posture angle of the host vehicle.   The information indicating the situation around the host vehicle includes the position of an object existing around the host vehicle, information indicating whether the object is moving or stationary, map information, and weather information around the host vehicle. The target object detection apparatus of any one of Claims 1-9 containing. Computer
An acquisition means for acquiring a captured image obtained by imaging the periphery of the host vehicle, information indicating a traveling state of the host vehicle, and information indicating a surrounding situation of the host vehicle,
Based on the information acquired by the acquisition means, a dangerous area in which the object may exist, and a dangerous area estimation means for estimating an area risk indicating the degree of danger in the dangerous area;
Based on the positional relationship between the dangerous area and the host vehicle and the posture of the object assumed to be in the dangerous area, the posture-specific risk level indicating the degree of danger for each posture is estimated. Risk estimation means,
Based on the risk level of the area and the risk level for each posture, a total risk level is estimated for each posture of the target object that is assumed to be present in the risk area, and In order to identify the target object in advance, selecting the dangerous area in descending order, extracting the window image from the area on the captured image corresponding to the selected dangerous area, and the window image extracted by the extracting means An image in which the window image represents an object based on an identification model selected according to the attitude of the object assumed to exist in the selected dangerous area from the identification model generated for each attitude of the object An object detection program for functioning as an identification means for identifying whether or not.   An object detection program for causing a computer to function as each means constituting the object detection device according to any one of claims 1 to 10.

Images