JP2008299787A - Vehicle detector - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To detect an object ahead of one's own vehicle by a plurality of on-vehicle sensors, to identify whether the object is a vehicle, to output the reliability of a detection result, and to support driving based thereon. <P>SOLUTION: This vehicle detector of the present invention is introduced to have a means for measuring nonvehicle (peripheral object) likelihood, for example, white line likelihood and road likelihood, in addition to a means for measuring vehicle likelihood, and does not determine a preceding vehicle position independently by the respective means, but detects the preceding vehicle position, based on an integrated scale integrated with the "vehicle likelihood" and the "nonvehicle likelihood". The reliability of detection is also calculated based on a distribution of a photographed environment and the integrated vehicle likelihood, so as to execute identification for three patterns of the "vehicle", "nonvehicle" and "difficulty in identification". <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an apparatus that detects an object ahead of a host vehicle by using a plurality of in-vehicle sensors, identifies whether the object is a car, and outputs the reliability of the detection result.

  It is difficult to accurately determine the position of the front vehicle with a single sensor. For this reason, there is a method for obtaining the front vehicle position with high accuracy by comprehensively judging the detection results of the plurality of detection means. Using the vehicle-likeness measuring means for measuring the vehicle-likeness ahead by image processing of millimeter wave radar output and camera video, the vehicle-likeness measuring means alone performs front-vehicle detection using the vehicle-likeness measuring means. Some integration means integrate the detection results obtained by the respective means to determine the front vehicle position, and control the movement of the vehicle by the vehicle control device according to the position information (for example, Patent Document 1).

Japanese Patent Laying-Open No. 2005-165421 (FIG. 1)

  However, in this method, (1) only “car-likeness” is used as information to be integrated, (2) each vehicle-likeness measuring means alone determines the front vehicle position and then integrates, (3) multiple The correlation between vehicle quality measurement means is not taken into account (conversely, integration is performed by integrating probability distributions generated from detection results on the premise of being independent), and (4) previous vehicle inspection "Car" is a mechanism that determines whether the reliability judgment of the knowledge result is "car" or "other than that (when it cannot be judged due to the influence of the shooting environment, or when there is something other than a car ahead)" The problem is that it is not a mechanism for determining “non-car” or “unidentifiable”.

  On the other hand, the present invention first improves the separation performance between the vehicle and the non-vehicle by measuring the vehicle-likeness and the non-vehicle-likeness by each measuring means. Next, the determination of the target position is not performed for each individual measuring unit, and the vehicle-likeness distribution by each measuring unit is integrated to create the vehicle-likeness distribution within the search range. In the preceding example, if the single detection means does not have high performance, the integrated target position becomes inaccurate, but in the present invention, high detection performance can be realized by delaying the determination after integration.

  The vehicle detection device of the present invention outputs a vehicle-likeness distribution that integrates the output of a vehicle-likeness / non-vehicleness-measuring means that measures vehicle-likeness and non-vehicle-likeness using a plurality of different sensors, and an output of the vehicle-likeness / non-vehicleness-likeness measuring means. Vehicle-likeness distribution output means, and vehicle position judging means for judging the position of the vehicle from the vehicle-likeness distribution.

  The vehicle detection device of the present invention outputs a vehicle-likeness distribution that integrates the output of a vehicle-likeness / non-vehicleness-measuring means that measures vehicle-likeness and non-vehicle-likeness using a plurality of different sensors, and an output of the vehicle-likeness / non-vehicleness-likeness measuring means. Since the vehicle-likeness distribution output means and the vehicle position judging means for judging the position of the vehicle from the vehicle-likeness distribution are provided, high detection performance can be realized.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a configuration of a vehicle detection device according to Embodiment 1 of the present invention. The vehicle-likeness measuring means 11 measures the vehicle-likeness using a plurality of different sensors such as a millimeter wave radar, a laser radar, and an image sensor. The signal obtained from the vehicle quality measurement means 11 is processed, and for example, the vehicle quality distribution output means 13 can determine the distribution of the vehicle quality in the feature space spanned by the two parameters of the left and right positions and the width of the front vehicle.

  For example, the distribution of the distance to the object by the millimeter wave radar, the speed distribution of the object by comparing the inter-vehicle distance by the millimeter wave radar and the own vehicle speed by the own vehicle speed sensor, and the vehicle likeness of the texture on the rear surface of the vehicle by the image sensor. Scales (such as the symmetry of the texture and the amount of vertical edge that hits the vehicle width) can be considered.

  For example, since the texture of the rear surface of a car is often left-right symmetric, it is more likely that the texture is more symmetrical. In addition, since there should be an edge between the left and right ends of the car and the background, it can be determined that there are left and right ends of the car at positions where many vertical edges are detected.

  In addition, a scale that measures the temporal and spatial continuity of such vehicle-likeness (estimated from the distribution of similarity to the previous detection of the texture on the rear surface of the vehicle, and the vehicle detection position and relative speed at the past frame) The vehicle existence probability distribution in the current frame to be detected, the distribution of the stability of the detected vehicle width, and the like) are also considered as 11 measures. If the same vehicle, the position does not move greatly and the speed does not change within a minute time, so there is a high possibility that there is a car around the position detected up to the previous frame, so the surrounding area is like a car. Is high. Further, since the texture on the rear surface of the vehicle should have a high correlation between frames, it is assumed that the vehicle quality at a position highly correlated with the texture of the vehicle position in the front frame is high. Also, since the vehicle width should not change, the vehicle width detected in the front frame is assumed to be a vehicle-like width.

  The non-car-likeness measuring means 12 is a means for measuring peripheral objects other than the vehicle, such as white line-likeness and road-likeness, by using a millimeter wave radar, a laser radar, an image sensor or the like, and obtaining the distribution. For example, if the white line position can be detected from the camera image by the white line detection algorithm, the white line near the white line position is high. Also, since the average brightness (or median brightness) of the area below the horizon of the camera image should be close to the brightness value of the road, the difference from that brightness is small, and the area with less texture is more likely to be a road Become.

  The non-vehicle-likeness measuring unit 12 may be a sensor used as the vehicle-likeness measuring unit 11. What is important here is to measure the vehicle-likeness and the non-vehicle-likeness using a plurality of different sensors and sensing techniques based on different methods.

  The vehicle-likeness distribution output unit 13 includes a plurality of vehicle-likeness distribution and non-vehicle-likeness measuring units 12 output by the vehicle-likeness measuring unit 11 in a feature space spanned by two parameters of the left and right positions and the width of the front vehicle. This is means for integrating a plurality of non-vehicle-like distributions output by each to generate one vehicle-likeness distribution.

  As an integration method in the vehicle-likeness distribution output unit 13, for example, there is a method of integrating the vehicle-likeness by subtracting the weighted linear sum of the non-vehicle-like index from the weighted linear sum of the vehicle-likeness index.

  In addition, by considering each vehicle-likeness and non-vehiclelikeness index as an axis that constitutes a multidimensional space, the vehicle and non-vehicle identification plane in this multidimensional space is obtained, and the vehicle is determined by the distance from that plane. A method to integrate the uniqueness can be considered. The discriminant plane is obtained by using linear discriminant analysis or using a support vector machine (SVM). It is also possible to obtain an identification surface in a high-order space using a nonlinear SVM using a kernel function, and to consider the distance from the identification surface as a highly accurate vehicle.

  The front vehicle detection means 14 searches for and outputs a car-like point from the vehicle-likeness distribution that is the output of the vehicle-likeness distribution output means 13 in the feature space spanned by the two parameters of the left and right position and width of the front car. A point that is maximum in terms of space in the distribution of carness is output. By analyzing the vehicle-likeness distribution by the front vehicle detection means 14, the vehicle position determination means 18 can determine the position of the vehicle. In response to this determination result, the driving support device 19 can assist in driving the vehicle. For this reason, a vehicle detection apparatus is normally provided in the own vehicle.

  In the drawings, the same reference numerals denote the same or corresponding parts, and this is common throughout the entire specification. Further, the description of the constituent elements appearing in the whole specification is merely an example and is not limited to these descriptions.

  When only the “car likeness” is used as information to be integrated, that is, when the non-car likeness measuring means 12 is not used, an area surrounded by a straight edge such as a white line or a power pole or a large car such as a truck Some parts tend to return a high value in any vehicle quality index. For example, when considering the symmetry of the back surface texture as an image characteristic obtained from camera images as an index of car likeness, the distribution of cars and non-cars overlaps greatly, and the combination of only the car likeness index increases the non-car and car. It may be difficult to separate accurately.

  On the other hand, in the present invention, by using the non-vehicle-likeness measuring means 12 in combination, it is possible to screen out objects such as white lines and utility poles that are highly likely to be erroneously detected. Indices such as white lines and utility poles are used to detect cases such as areas surrounded by straight edges such as white lines and utility poles and parts of large vehicles such as trucks that are falsely detected only by the indications of the car. By making a comprehensive judgment, the degree of separation can be improved.

  Moreover, it differs from what integrates, after determining a vehicle position by each vehicle-likeness measurement means independently. For example, in the configuration of the preceding example shown above, the detection performance of each vehicle-likeness measurement means alone is low, so the front vehicle position is determined by each means independently of the use of a plurality of means. Is integrated. It is presumed that the single measuring means has poor detection performance and often has a higher vehicle quality than the vehicle at positions other than the correct vehicle position.

  As described above, the vehicle detection device uses a plurality of different sensors to measure the vehicle-likeness and the non-vehicle-likeness, and the vehicle-likeness distribution that integrates the output of the vehicle-likeness / non-vehicleness-likeness measuring means and the vehicle-likeness / non-vehicleness-likeness measuring means. Since the vehicle-likeness distribution output means for outputting the vehicle position and the vehicle position determination means for judging the position of the vehicle from the vehicle-likeness distribution are provided, high detection performance can be realized.

Embodiment 2. FIG.
FIG. 2 is a block diagram showing the configuration of the vehicle detection device according to Embodiment 2 of the present invention. The following description will focus on differences from the first embodiment.

  The distribution reliability measuring means 15 is a means for calculating the likelihood of the front vehicle position, that is, the distribution reliability of the vehicle likelihood distribution from the vehicle likelihood distribution obtained by the vehicle likelihood distribution output means 13. For example, the value of the vehicle likeness at the detection position itself, the sharpness of the distribution of the vehicle likeness, the difference from the second peak value around the detection position, and an index indicating whether the distribution is ridged. The greater the difference in the vehicle quality from the vicinity of the detection position is considered to be the detection with higher reliability. Therefore, the higher the sharpness of the distribution of vehicle quality, the higher the reliability, and the greater the difference from the second peak value. Is considered highly reliable.

  On the other hand, the environmental reliability measurement means 16 is a means for outputting an index indicating whether the environment measured by the vehicle quality measurement means 11 is easy or difficult to detect the vehicle, and the environmental reliability of the environment at the time of sensor measurement. It measures. For example, the brightness of the sky at the time of photographing, the complexity of the background, the edge amount of the car area, the contrast of the car area, and the like can be mentioned.

  For example, when an image sensor is used as the vehicle-likeness measuring unit 11, if the brightness of the sky at the time of shooting is dark, the possibility of nighttime increases, and if it is too bright, the possibility of a backlight scene during the day increases. It can be determined that the brightness is the most reliable. As for the complexity of the background, if the background is complex, there is a high possibility that there are many false detection targets such as urban areas. In addition, when the edge amount of the car area is small or the contrast is low, it can be determined that it is difficult to detect because there is a high possibility of being backlit at the time of shooting or having a few features in the detection target.

  The detection reliability calculation unit 17 is a unit that outputs the reliability of the detection result by integrating the indexes obtained by the distribution reliability measurement unit 15 and the environment reliability measurement unit 16. An index of detection reliability can be obtained by calculating a product or a sum of a plurality of indices.

  The vehicle position determination unit 18 uses the detection result of the front vehicle detection unit 14, the vehicle likelihood value of the vehicle likelihood distribution output unit 13, and the detection reliability of the detection reliability calculation unit 17, so that the vehicle, non-car, Identify three states that are difficult to determine.

  FIG. 3 is a schematic diagram showing the relationship between the vehicle quality and the reliability. The distinction between cars and non-cars is based on the axis of car-likeness and reliability. When the reliability is high, “car” and “non-car” are clearly identified, and when the reliability is low, the area of “difficult to judge” is Take a mechanism to grow.

  The driving support device 19 performs driving support according to the identification result obtained by the vehicle position determination means 18. For example, in the case of a collision-reducing brake system, if a vehicle is detected, the risk of collision is derived from the amount of change in the left-right position, width, and inter-vehicle distance, and brake control is performed at an early timing as necessary. Take control. If it is identified as a non-vehicle, it is determined that there is no vehicle at the detected location (in short, erroneous detection), and brake control is not performed. If it is determined that the judgment is difficult, the conventional control is performed using only the output of the millimeter wave radar, for example, according to the operation of the conventional safety system.

  In addition, based on the above-mentioned collision risk, avoidance braking by steering angle control, seat belt pull-in control, warning sound, warning display on the display device and front vehicle position display, etc. Driving assistance such as can be considered.

FIG. 4 is a schematic diagram for explaining the likelihood. Here, in the case of making a comprehensive determination using two indexes having a vehicle-likeness (likelihood) distribution as shown in FIG. 4 (a), in the preceding example, as shown in FIG. A probable position is determined (x ₁ and x _{2 in} this example), and the final front vehicle position is determined based on this position. In this example, the possibility of erroneous detection the final preceding vehicle located at a position between x ₁ and x ₂ is increased. On the other hand, in the present invention, as shown in (c), the position where the integrated likelihood becomes maximum can be detected at x ₃ .

  Furthermore, it is different from the case where the correlation between the plurality of vehicle quality measurement means is not considered. In other words, it is different from the case of simply integrating the integration of probability distributions generated from the detection results under the premise of being independent. Even if the means are different, it is difficult to imagine a case where the indicators that measure the same car quality are not correlated and independent, and in an extreme case, these sensors are not useful when two or more of the same millimeter wave radars are installed side by side. The position of the front vehicle detected by is completely coincident. In such a case, if the probability distribution is integrated by integration, even if the detection results from other camera images point to different locations, the detection results from the millimeter wave radar are accumulated multiple times. The position pointed to by the millimeter wave radar is the integration result. On the other hand, in the present invention, since the determination is made at the position where the integrated likelihood is the maximum, high detection performance can be realized unlike simple integration.

  In addition, a mechanism that simply determines whether the reliability judgment of the detection result of the front vehicle is “car” or “other than that (when it cannot be judged due to the influence of the shooting environment, or when there is something other than the car ahead)” Is different. Since it is impossible to detect the front vehicle position with 100% accuracy, when using the calculated front vehicle position in a safety system, a mechanism is required to determine whether the detection result is detected with high reliability. It becomes. In the preceding example shown above, as an index of reliability, either an index based on statistical data or an evaluation value itself at the time of actual detection is defined as the reliability, and confidence depends on the magnitude of that one parameter. Because it is determined whether or not it can be said that a car has been detected with the camera, and if it cannot be said that it is a car, it is difficult to detect the object itself due to the influence of the shooting environment, etc. It is impossible to determine whether the reliability is low. On the other hand, the present invention distinguishes from “car”, “non-car”, and “difficult to identify” by realizing a mechanism for adjusting sensitivity using two parameters of reliability and car-likeness value at the time of detection. And high detection performance can be realized.

  In the present invention, in addition to (A) means for measuring vehicle-likeness, means for measuring non-car (peripheral object) -likeness, for example, white line-likeness, road-likeness, and the like are introduced. Further, (B) the front vehicle position is detected from the integrated scales by integrating the “scale of vehicle-likeness and non-vehicle-likeness” without determining the front vehicle position by each measuring means alone. In addition, (C) “Car”, “Non-car”, “Difficult to identify” by taking into account two parameters of detection reliability and the value of car-likeness itself at the time of detection based on the shooting environment and the state of integrated car-likeness distribution. Is identified into three patterns.

  As described above, the environmental reliability measuring means for measuring the environmental reliability of the environment at the time of sensor measurement, and the detection reliability calculating means for calculating the detection reliability from the distribution reliability and the environmental reliability of the vehicle-likeness distribution. In addition, the vehicle position determination means uses the detection reliability to identify the three states of car, non-car, and difficult to determine, so that higher detection performance can be realized.

It is a block diagram which shows the structure of the vehicle detection apparatus by Embodiment 1 of this invention. It is a block diagram which shows the structure of the vehicle detection apparatus by Embodiment 2 of this invention. It is a schematic diagram which shows the relationship between vehicle character and reliability. It is a schematic diagram for demonstrating likelihood.

Explanation of symbols

11 vehicle-likeness measurement means, 12 non-vehicle-likeness measurement means, 13 vehicle-likeness distribution output means, 14 front vehicle detection means, 15 distribution reliability measurement means, 16 environmental reliability measurement means, 17 detection reliability calculation means, 18 vehicle position Judging means, 19 Driving support device.

Claims (2)

Vehicle-like / non-vehicle-likeness measuring means for measuring vehicle-likeness and non-vehicle-likeness using a plurality of different sensors;
Vehicle-likeness distribution output means for outputting a vehicle-likeness distribution that integrates the output of the vehicle-likeness non-vehicle-likeness measuring means;
A vehicle detection device comprising vehicle position determination means for determining a vehicle position from the vehicle-likeness distribution. Environmental reliability measurement means for measuring the environmental reliability of the environment at the time of sensor measurement;
A detection reliability calculating means for calculating a detection reliability from the distribution reliability of the vehicle-likeness distribution and the environmental reliability;
2. The vehicle detection device according to claim 1, wherein the vehicle position determination unit identifies three states of a vehicle, a non-vehicle, and a determination difficulty using the detection reliability.

Images