JP2012118683A - Pedestrian recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately recognize a pedestrian by a configuration in which the number of patterns to be collated is small and costs are reduced.SOLUTION: By part specifying means of an image recognition processing part 4, a part in the case of defining that an obstacle around a present vehicle 1 detected by a distance measuring radar 2 and a monocular camera 3 is a pedestrian 6 is specified. By posture determination means of the image recognition processing part 4, from the collation of the predicted posture of the part of the pedestrian 6 read from a storage part 5 by predicting a posture change of the specified part and the posture of the part specified by the part specifying means, whether or not the movement of the specified part matches with the movement of a pedestrian is determined. By recognition means of the image recognition processing part 4, whether or not the obstacle is a pedestrian is determined on the basis of a determination result of the posture determination means and the pedestrian is recognized.

Description

  The present invention relates to a pedestrian recognition apparatus that recognizes a pedestrian by pattern matching, and more particularly, to an improvement in recognition accuracy with reduced cost.

  Conventionally, in the field of automobiles, it is important for collision avoidance and the like to recognize whether or not an obstacle around the host vehicle, such as in front of the host vehicle, detected by radar exploration or camera photography is a pedestrian.

  And, in order to recognize as accurately as possible whether or not the captured obstacle is a pedestrian by pattern matching of the captured image of the camera, conventionally, the pattern of the entire obstacle when the obstacle is assumed to be a pedestrian, Instead of memorizing, the pattern for each part such as the pedestrian's head, torso, and leg is individually stored in the memory means, and the identified part of the detected obstacle and each part stored. It has been proposed to recognize the obstacle by determining whether or not the obstacle is a pedestrian (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2009-25959

  In the pedestrian recognition described in Patent Document 1, pedestrians are sequentially collated with patterns of all postures (including directions of the pedestrians with respect to the own vehicle) stored in the storage unit corresponding to the specified part. Therefore, a huge amount of pattern matching or the like is necessary, and their processing load becomes extremely large. Therefore, an expensive image processing apparatus having a high processing speed is required, and there is a problem that the cost becomes high.

  An object of the present invention is to recognize a pedestrian with high accuracy by a configuration in which the number of patterns to be compared is small and the cost is reduced.

  In order to achieve the above-described object, the pedestrian recognition apparatus of the present invention specifies a region when the obstacle around the host vehicle detected by the detection unit is a pedestrian, and specifies the specified region and storage unit. Is a pedestrian recognition device that recognizes a pedestrian by determining whether or not the obstacle is a pedestrian based on the result of the comparison, and is detected by the detection means A part specifying unit that specifies a part when the obstacle is a pedestrian, and a prediction of the part of the pedestrian read from the storage unit by predicting a posture change of the part specified by the part specifying unit A posture determination unit that determines whether or not the movement of the identified part matches the movement of the pedestrian based on a comparison between the determined posture and the posture of the part specified by the part specifying unit; The obstacle from the judgment result Is characterized by comprising a recognition means for recognizing the pedestrian by determining whether a pedestrian (claim 1).

  According to the first aspect of the present invention, when a part when the obstacle around the host vehicle detected by the detecting unit is a pedestrian is specified by the part specifying unit, the part is determined by the posture determining unit. If it is a part of a pedestrian, what kind of movement will be performed next (what kind of posture will be taken) is predicted, and there are few from the collation with the predicted posture and the posture of the part specified by the part specifying means Even in the case of pattern matching, it can be determined whether or not the movement of the part specified with a high pattern matching probability matches the movement of the pedestrian. Then, from the determination result of the posture determination means, it is possible to accurately determine whether the obstacle is a pedestrian by the recognition means. In this case, the processing load for pattern matching and the like is small, and there is no need for a high-performance and expensive image processing device with a high processing speed. Can recognize.

It is a block diagram of one embodiment of a pedestrian recognition device of the present invention. It is explanatory drawing of the site | part which the pedestrian recognition apparatus of FIG. 1 specifies. It is explanatory drawing of the example of an attitude | position pattern of the pedestrian recognition apparatus of FIG. It is explanatory drawing of the example of a state transition of the pedestrian recognition apparatus of FIG. It is a flowchart explaining the example of a process sequence of the pedestrian recognition apparatus of FIG.

  An embodiment of the present invention will be described with reference to FIGS.

  FIG. 1 shows a pedestrian recognition apparatus of the present embodiment mounted on a host vehicle 1. The host vehicle 1 includes a ranging radar 2 such as a laser radar or an ultrasonic radar that searches the front of the host vehicle 1, and the host vehicle. A monocular camera 3 that captures the front of 1 is provided as an obstacle detection means around the host vehicle of the present invention. In the case of this embodiment, it is difficult to grasp the presence of an obstacle ahead by using only an inexpensive image of the monocular camera 3, so the combination of the ranging radar 2 and the monocular camera 3 is used as the detection means. When a so-called stereo camera is provided instead of the camera, the detection means is formed only by this camera.

  Then, the result of the search by the ranging radar 2 and the captured image of each frame of the monocular camera 3 are input to the image recognition processing unit 4 formed by a microcomputer or the like. The image recognition processing unit 4 detects a range of obstacles ahead from the search result of the ranging radar 2, extracts the range of the obstacles in the captured image of the monocular camera 3 as an ROI (region of interest), and extracts it. The area image is an obstacle image.

  Further, the image recognition processing unit 4 basically specifies a part when the obstacle image is a pedestrian by a pattern matching method, and specifies the specified part and the storage unit 5 as a storage unit of the present invention. The pedestrian part pattern stored in the pedestrian is collated, and the pedestrian is recognized by determining whether the obstacle is a pedestrian based on the collation result. Forming a recognition means.

  By the way, the pedestrian can be regarded as a set of several parts from the feature of the posture change.

  FIG. 2 shows an example of the above-mentioned part of the pedestrian 6. In this embodiment, the pedestrian 6 is divided into a trunk portion 61, an arm portion (including the hand) 62 having a posture change, and a leg portion (including the toe). ) 63, and the basic pattern of the torso 61, the arm (including the hand) 62, and the leg (including the toe) 63, and information on how to move and the range of movement (angle) are stored in the storage 5. Are stored in advance in the pedestrian basic dictionary 51.

  Next, the part pattern of the arm part 62 and the leg part 63 not only varies depending on the posture of the pedestrian 6, but also changes based on the movement of the pedestrian predicted from the posture. The next posture can be predicted.

  Therefore, for the arm part 62 and the leg part 63, various posture patterns (posture patterns) are stored in advance in the posture dictionary 52 of the storage unit 5 with the transition information of the movement order as their part patterns. To do.

  FIGS. 3A to 3G show examples of posture patterns P1 to P7 classified as posture 1 to posture 7 of the arm portion 62. Each posture pattern P1 to P7 is the next predicted from the movement of the arm. Transition information (transition flag) with a probability of priority order to one or a plurality of posture patterns is attached. The probability of the priority order of the next one or more posture patterns expected here will be described. For example, the current posture pattern of the arm portion 62 of the pedestrian 6 is the posture 1, and the next one or more predicted. If the posture pattern becomes posture 1 or posture 2, the probability of high priority order is assigned to posture 1 and posture 2. For the legs 63, posture patterns with similar priority order probabilities are stored in the posture dictionary 52.

  FIG. 4 shows an example of a change in the posture of a part of the posture based on the transition flag, and if it is a pedestrian 6, it is easy to change from posture patterns P1 to P7 to a posture pattern with high probability. If the posture is changed such that there is no portion determined to be the part 63 or a transition with an extremely low probability, it can be understood that it is not the pedestrian 6.

  In the case of the present embodiment, the arm portion 62 or the leg portion of the obstacle is based on a comparison between the posture pattern of the portion determined to be the arm portion 62 or the leg portion 63 and a pattern that is equal to or higher than the set probability of the expected posture transition. 63 posture changes are estimated, and depending on whether or not the expected posture transitions of the arm part 62 and the leg part 63 of the obstacle are not less than the set probability in a set time, the posture change is natural as a pedestrian. Whether the obstacle is a pedestrian 6 or not is also estimated, and the walking vehicle 6 is recognized.

  FIG. 5 shows an example of the procedure of the pedestrian recognition process executed by the part specifying means, posture determining means, and recognizing means of the image recognition processing unit 4. First, the body part and arm part of the obstacle image are shown by the part specifying means. The candidate leg portion is specified from the pattern, image position, and the like (step S1). Next, pattern matching is performed between the pattern of the candidate part of the specified body part and the pattern of the body part 61 of the basic dictionary 51 (step S2), and a pattern of the body part 61 that matches at or above the threshold is searched. If a matching body part 61 is found above the threshold value, pattern matching is performed between the pattern of the identified arm and leg candidate parts and the pattern of the arm part 62 and leg part 63 of the basic dictionary 51 (step S3). The pattern of the arm part 62 and the leg part 63 that match at a threshold value or more is searched. If the matching arm part 62 and leg part 63 are found above the threshold value, the matched trunk part 61, arm part 62 and leg part 63 are combined and integrated (step S4).

  Next, the state transition (time change) of the posture of the arm unit 62 and the leg unit 63 in the integrated state is detected by the estimation unit, and the transition probability of the posture order is tracked by pattern matching with reference to the posture dictionary 52. The movement of the arm part 62 and the leg part 63 by detecting whether or not the probability of the natural movement of the pedestrian exceeds a threshold value based on whether or not the posture changes with a predetermined probability within the set time. Is a change that matches the movement of the pedestrian 6 (step S5), and based on the estimation result, it is determined whether the obstacle is the pedestrian 6 by the recognition means. The obstacle is recognized as a pedestrian 6 (step S6).

  In this case, the pattern of the pedestrian 6 is divided into the body part 61, the arm part 62, and the leg part 63, and those basic patterns are stored separately in the basic dictionary 51, and the probability of matching is high from individual pattern matching. By combining the patterns of the body part 61, the arm part 62, and the leg part 63 to perform basic pattern matching and detection of changes in posture patterns, the pattern is not divided into the body part 61, arm part 62, and leg part 63. Compared to the case of matching, the number of patterns to be stored is extremely small, and the processing load of the image recognition processing unit 4 is greatly reduced.

  Furthermore, regarding the arm part 62 and the leg part 63, the movement (posture change) is estimated by referring to the posture pattern of the posture dictionary 62 from the state transition probability to determine whether or not it is a natural movement that seems to be a pedestrian 6, and extremely high collation. It can be recognized with probability that the obstacle is a pedestrian 6.

  Therefore, in the case of this embodiment, the pedestrian 6 can be recognized with high accuracy by the configuration in which the number of patterns to be compared is small and the cost is reduced.

  The present invention is not limited to the above-described embodiment, and various modifications other than those described above can be made without departing from the spirit thereof. Etc. are not limited to the above embodiment.

  Further, the state transition of the posture change and the processing procedure are not limited to the above embodiment. Furthermore, it goes without saying that obstacles other than the front such as the rear of the host vehicle 1 can be recognized in the same manner.

  And the recognition result of this invention can be utilized for various driving assistance control.

DESCRIPTION OF SYMBOLS 1 Own vehicle 2 Ranging radar 3 Monocular camera 4 Image recognition process part 5 Memory | storage part 6 Pedestrian 61 Body part 62 Arm part 63 Leg part

Claims (1)

Based on the result of the collation, the part when the obstacle around the own vehicle detected by the detection means is a pedestrian is identified, the identified part and the part pattern of the pedestrian stored in the storage means are collated , A pedestrian recognition device for recognizing a pedestrian by determining whether the obstacle is a pedestrian,
A part specifying means for specifying a part when the obstacle detected by the detecting means is a pedestrian;
The identification from the collation between the predicted posture of the pedestrian of the pedestrian read out from the storage means by predicting the posture change of the part specified by the part specifying means and the posture of the part specified by the part specifying means Posture determination means for determining whether or not the movement of the selected part matches the movement of the pedestrian,
A pedestrian recognition apparatus comprising: a recognition unit configured to determine whether the obstacle is a pedestrian based on a determination result of the posture determination unit and to recognize the pedestrian.

Images