JP5212004B2 - Vehicle tracking device and vehicle tracking method - Google Patents

Description

  The present invention relates to a vehicle tracking device and a vehicle tracking method, and more particularly to a vehicle tracking system for tracking a vehicle present in an image captured by a camera.

  “Safety and security” is a keyword in recent automobile development. As the Japanese society ages, it is inevitable that the elderly population of drivers will increase in the future, but support safe and reliable driving for elderly drivers whose judgment, attention, and physical strength have declined. To do is highly desirable for the automotive industry. Safety and security for automobiles can be thought of as avoiding the occurrence of traffic accidents or minimizing the damage if they occur.

  Although the number of fatalities in traffic accidents has been steadily decreasing from 9,640 in 1997 to 6,352 in 2006 due to improvements in the performance of automobiles and revisions to the Road Traffic Law, It is almost flat, and further improvements are demanded regarding the safety of automobiles. For example, a vehicle collision avoidance system using a millimeter-wave radar and a night road by displaying a far-infrared image on a car navigation screen as a technology for recognizing the surrounding situation of a vehicle necessary for avoiding a collision between a vehicle and a person or an object An image clarification system and a lane departure warning system by recognizing a white line using a front camera have been put into practical use.

  These are technologies in which processing is completed with one vehicle, but there are vehicle-vehicle cooperation technology and road-vehicle cooperation technology as approaches for realizing higher-level situation recognition. With vehicle-to-vehicle cooperation, each vehicle traveling on the road communicates wirelessly with each other, and the surrounding situation of the own vehicle is provided to other vehicles, so that it is possible to acquire the surrounding situation that is difficult to grasp only with the own vehicle. Become. By realizing vehicle-vehicle cooperation, for example, it is expected to avoid a collision accident at the time of encounter due to another vehicle popping out from the blind spot.

  In road-vehicle cooperation, the running state of each vehicle is monitored using a sensor installed beside the road, and the running information is distributed to each vehicle. By realizing road-vehicle cooperation, for example, it becomes possible to detect and notify the end of a traffic jam. A camera may be used as one of the sensors used in road-vehicle cooperation. By detecting individual vehicles from the road images acquired by the cameras and tracking their travel routes, the vehicle goes straight with a right turn car at the intersection. It is also possible to construct a car collision prevention warning system. Advantages of using a camera as a sensor include that it is possible to monitor a wide range with simple setting and low cost, and that an individual vehicle can be measured at an arbitrary position on the road.

  On the other hand, as a major problem, in the case of so-called occlusion where a vehicle of interest is shielded by another vehicle and part of the vehicle of interest is observed missing on the image, the vehicle of interest is likely to be undetected. The point that becomes. In particular, in the vicinity of an intersection where road-vehicle cooperation technology is strongly desired, the vehicle stops when waiting for a signal, and occlusion occurs frequently on the image. Therefore, in vehicle-vehicle cooperation using an image sensor, a vehicle tracking technique that is robust against occlusion is essential. The following are known as conventional techniques for detecting a vehicle from a sensor installed on the roadside and tracking the travel route.

  With reference to Patent Document 1, templates are prepared for each portion constituting the vehicle of interest, and individual vehicles are detected by matching these templates with objects in the input image. At that time, by performing matching only in the vicinity of the predicted position one hour ahead obtained by applying a Kalman Filter to the vehicle position, the vehicle tracking is improved in accuracy.

  In addition, referring to Patent Document 2, in a situation where the position and speed information of the preceding vehicle is required, such as an automatic following traveling function using a millimeter wave radar, the preceding vehicle cannot be detected as a vehicle due to a decrease in the speed of the preceding vehicle. In this case, the position of the preceding vehicle at the current time is estimated from the information on the speed and position of the preceding vehicle so far.

  Further, referring to Patent Document 3, in order to improve the estimation accuracy of the traffic state in the road section where the vehicle detection sensors are not densely installed on the road side, as shown in FIG. "Cells" divided into equal intervals (hereinafter referred to as cell images). The traffic flow is regarded as a compressible fluid, and the density of vehicles in each cell, the spatial average speed, etc. are used as model internal states, and the Kalman Filter is used for the estimation.

  Further, a traffic flow simulation model disclosed in Non-Patent Document 1 is well known that uses a Cell Automaton (CA). In this CA, each cell of the cell image is divided into a vehicle presence / absence cell. When the vehicle presence cell of interest meets the update condition, the vehicle presence cell of interest is moved. Examples of the update condition include a rule 184 and a Nagel-Schreckenberg algorithm.

  Now, assuming that the left to right of the cell image is the vehicle traveling direction, the rule 184 pays attention to each vehicle existing cell at the current time as shown in FIG. In the case of an absent cell, the target vehicle presence cell is moved to the cell on the right at the next time, and in the case of a vehicle presence cell, the position of the target vehicle presence cell is not moved.

The Nagel-Schreckenberg algorithm considers actual vehicle behavior (acceleration and deceleration) in the above-mentioned rule 184, and the contents can be summarized as follows.
(1) Let v (cell unit) be the speed of the vehicle of interest at the current time.
(2) Acceleration: If v <vmax and the distance to the vehicle ahead is v + 1 or more, the speed is increased by 1 (v → v + 1).
(3) Deceleration: When the vehicle of interest is in cell i and the vehicle ahead is in cell i + j (j <v), the speed is reduced to j−1 (v → j−1).
(4) Randomization: When v> 0, the speed is decreased by one with probability p (v → v−1).
(5) Move: Advance v cells.

  As can be seen from the Nagel-Schreckenberg algorithm, the traffic flow simulation model using Cell Automaton enables simulation that eliminates the situation that is impossible in reality, such as overtaking vehicles on the same lane.

  As another technique, there is one disclosed in Patent Document 4. In Patent Document 4, in order to suppress the influence of occlusion in the detection and tracking of a plurality of objects as much as possible, a plurality of cameras are installed around the observation region, and aperture information from each direction of each object obtained from these cameras is obtained. Are integrated to estimate the internal state (position and direction and the amount of change thereof) of each object that is probabilistically plausible.

JP 2001-357403 A JP 2001-043499 A Japanese Patent Application Laid-Open No. 2004-077842 JP 2004-220292 A Sugiyama Yuuki, “Physics of Traffic Flow”, Nagare 22, 2003, p95-p108

として、ろ過処理の結果得られる隠れ状態が従う正規分布の平均ベクトルと分散共分散行列とを、それぞれ、

とする。 In the technique of Patent Document 1, a conventional Kalman filter is used for tracking an object. The Kalman filter is a method for estimating a hidden state using given observation data, and includes a filtering process for calculating a hidden state at the current time and a prediction process for calculating a hidden state one hour ahead. The observation data at time n is

As a mean vector and a variance covariance matrix of normal distributions obeyed by the hidden state obtained as a result of filtration,

And

とすると、カルマンフィルタでは異なる時刻の隠れ状態間でマルコフ性が、隠れ変数と観測データの間では線形の演算がそれぞれ成り立つと見なして、

と、それぞれ定式化する。 Similarly, the mean vector of the normal distribution and the variance-covariance matrix that the hidden state obtained as a result of the prediction process follows,

Then, in the Kalman filter, it is assumed that Markov property is established between hidden states at different times, and linear operations are established between hidden variables and observed data.

And formulate each.

は、それぞれ行列分散がＱ，Ｒのホワイトノイズである。このときろ過処理、予測処理は、それぞれ、

で逐次計算される。 Equation (1) is called a system equation, and Equation (2) is called an observation equation. Matrix F is a state transition matrix, H is an observation matrix,

Are white noises with matrix variances of Q and R, respectively. At this time, the filtration process and the prediction process are respectively

Is calculated sequentially.

で計算される。例えば、隠れ状態を現時刻および１時刻前の真の車両位置

を用いて、

と定義し、隣接した３時刻間での車両運動が等速直線運動で近似できるとすると、状態遷移行列と観測行列は、

の形で記述できる。 Note that the matrix K in equations (3) and (4) is called Kalman gain,

Calculated by For example, if the hidden state is the current vehicle time and the true vehicle position one hour before

Using,

If the vehicle motion between three adjacent times can be approximated by constant velocity linear motion, the state transition matrix and the observation matrix are

Can be described in the form of

  As can be seen from the equations (1) to (10), the Kalman filter is a process in which the accuracy is guaranteed when it is assumed that the observation data is given every time. When the occlusion of the vehicle of interest is significant, a detection failure occurs, and corresponding observation data does not exist (hereinafter abbreviated as a missing measurement state). When there is no observation data, the predicted position calculated one time ago is generally used as the filtering position at the current time. However, the longer the missing state, the greater the deviation from the actual vehicle behavior. As a result, tracking accuracy decreases.

  In the technique of Patent Document 2, when the speed of the preceding vehicle is equal to or lower than a preset threshold value, the absolute position and the absolute speed of the host vehicle, the relative speed between the position of the host vehicle accumulated so far, and the amount of change in the relative speed From the maximum likelihood estimation method, the absolute position and the absolute speed of the preceding vehicle at the current time are obtained. As in the technique of Patent Document 1, the actual behavior of the preceding vehicle becomes longer as the low-speed traveling state below the threshold becomes longer. The divergence increases.

  In the technique of Patent Document 3, the travel speed information of a vehicle that transmits probe information to the traffic volume simulation model and the information on the traffic flow and speed from the vehicle detection sensor installed on the road side are input, and the Kalman filter Although the state variables are estimated, the estimation accuracy deteriorates when the missing state of these pieces of information becomes long, as in the above-mentioned Patent Documents 1 and 2.

  In addition, the techniques of Patent Documents 1 to 3 are based on the premise that the same vehicle number is assigned to the same vehicle. If this assignment is not correct, the process fails. If the Kalman filter is used as usual in situations where the missing state is likely to be prolonged, such as when detecting a vehicle from a camera image where occlusion occurs, the actual vehicle will be deviated from the difference between the estimated state and the actual behavior of the preceding vehicle. And the vehicle number are shifted, and the premise of the same vehicle / the same vehicle number is not satisfied.

  In the technique of Patent Literature 4, the object number i and the internal state x are associated with each other one-to-one, and the internal state is expressed by a probability distribution. Therefore, in the update calculation of the probability distribution at each time, there is a possibility that the probability distribution of the object i and the object j may be switched due to the influence of occlusion, and in this case, the original object number correspondence is restored. Therefore, there is a problem that reliability is poor.

  SUMMARY OF THE INVENTION An object of the present invention is to provide a vehicle tracking device and a vehicle capable of highly accurate and highly efficient vehicle tracking that are robust against occlusion while preventing the temporal continuity between a tracking target vehicle object and a vehicle number from being interrupted. It is to provide a tracking method and its program.

  The vehicle tracking device according to the present invention includes vehicle detection means for detecting individual vehicles as objects from road images, vehicle number assignment means for assigning a plurality of vehicle numbers to the detected individual objects, and the vehicle number assignment. A travel route tracking means for tracking the travel route based on each of the vehicle numbers assigned by the means, and a probability probability of each route is calculated as a likelihood from the tracking route information obtained by the route route tracking means. Route likelihood calculating means and vehicle number pruning means for deleting a vehicle number corresponding to a route having a smaller likelihood calculated by the route likelihood calculating means from among the vehicle numbers assigned by the vehicle number assigning means It is characterized by including.

  The vehicle tracking method according to the present invention includes a vehicle detection step for detecting individual vehicles as objects from a road image, a vehicle number assignment step for assigning a plurality of vehicle numbers to the detected individual objects, and the assigned A travel route tracking step for tracking a travel route based on each of the vehicle numbers, a route likelihood calculation step for calculating the probability of each route from the track route information as a likelihood, and the assigned vehicle number A vehicle number pruning step for deleting a vehicle number corresponding to a route having a small likelihood calculated by the route likelihood calculating step.

  A program according to the present invention is a program for executing a vehicle tracking method by a computer, and includes a process for detecting individual vehicles as objects from road images, and a plurality of vehicle numbers for the detected individual objects. A process of assigning, a process of tracking a travel route based on each of the assigned vehicle numbers, a process of calculating the probability of each route from the tracking route information as a likelihood, and the assigned vehicle number And a process of deleting a vehicle number corresponding to a route having a smaller likelihood calculated by the route likelihood calculating step.

  According to the present invention, when a missing state occurs in vehicle tracking using a road image, by assigning a plurality of vehicle numbers to the vehicle detection result, temporal continuity between the vehicle detection result and the vehicle number is increased. In addition to securing the vehicle number, the vehicle number having a low likelihood obtained as a result of the tracking process is deleted, so that the efficiency of the tracking process can be improved. In addition, according to the present invention, the accuracy of tracking processing can be improved by increasing the likelihood when observation data exists and decreasing the likelihood when observation data is present. is there.

  Hereinafter, the present invention will be described in detail with reference to the drawings. FIG. 1 is a schematic functional block diagram of an embodiment of the present invention. Referring to FIG. 1, a vehicle tracking device according to an embodiment of the present invention detects a vehicle from a road image and outputs it together with coordinates, and an individual vehicle that is a processing result of the vehicle detection unit 1. Vehicle number allocating unit 2 for allocating a plurality of vehicle numbers by regarding the vehicle as a vehicle object, and using the position information of each vehicle object obtained from vehicle detecting unit 1 and the vehicle number allocated by vehicle number allocating unit 2, From the travel route tracking unit 3 that tracks and records the travel route of the vehicle object corresponding to each vehicle number, and the tracking route information that is the output result of the travel route tracking unit 3, the probabilistic likelihood of each route is estimated. Of the vehicle numbers assigned by the route likelihood calculating unit 4 and the vehicle number assigning unit 2, the vehicle number corresponding to the route having the smaller likelihood calculated by the route likelihood calculating unit 4 And a vehicle number pruning unit 5 for removed from objects.

  Next, the overall operation of the present embodiment shown in FIG. 1 will be described. The vehicle detection unit 1 detects a vehicle from a road image obtained from a camera installed on the road side. The vehicle number assigning unit 2 assigns a vehicle number to a vehicle object that is detected by the vehicle detecting unit 1 and determined to start tracking the travel route. The operation of the vehicle number assigning unit 2 will be described with reference to FIG. 2 showing the tracking state of the vehicle object at time t = T.

  In FIG. 2, it is assumed that the vehicle object α is determined to be an object for starting tracking of the travel route. The vehicle number assigning unit 2 first calculates the position of the vehicle object that is closest to the front and rear of the vehicle object α among the tracking vehicle objects existing on the lane 2 on which the vehicle object α travels ( In FIG. 2, vehicle objects A and B are represented).

  Further, among the tracked vehicle objects existing on the lane 1 and the lane 3 adjacent to the vehicle object α, search forward and backward from the same position (dashed line in FIG. 2) as the vehicle object α in the vehicle traveling direction. Then, the position of the vehicle object closest to the base point is calculated (represented by vehicle objects C and D in FIG. 2). A new unused vehicle number is adopted and assigned to the vehicle object α in accordance with the vehicle numbers assigned to the vehicle objects being tracked.

  In the case of FIG. 2, since vehicle numbers {10, 21}, {18}, {33, 47}, {53} are assigned to the vehicle objects A to D, the unused vehicle numbers 80 and these are assigned. The seven vehicle numbers {80/10, 21, 18, 33, 47, 53} are assigned. In the vehicle number set, the right side of / represents the inherited vehicle number, and the left side represents the newly adopted vehicle number.

  The traveling route tracking unit 3 tracks the traveling route of each vehicle object using the vehicle number assigned by the vehicle number assigning unit 2 and records the result. The operation of the travel route tracking unit 3 will be described with reference to FIGS. FIG. 3 shows a branch of a route that occurs by assigning the vehicle number held by the vehicle objects A and B to the vehicle object α. As can be seen from FIG. 3, the tracking of the travel route by assigning the vehicle number 10 to the vehicle object α is performed on the assumption that the vehicle object α is a vehicle traveling on the second route of the vehicle number 10. Means.

  Considering similarly, the vehicle object α travels on the second route of the vehicle number 10, the second route of the vehicle number 18, the second route of the vehicle number 21, and the first route of the vehicle number 80. It represents a vehicle that can be held as a route. The position of the vehicle object α is calculated using the route having the highest likelihood among the plurality of routes that the vehicle object α can hold.

  FIG. 4 explains vehicle position estimation in a plurality of routes possessed by each vehicle number described in FIG. 3, and represents a travel route of a vehicle object to which the vehicle number 10 is assigned. From FIG. 2, there are vehicle objects α and A as the actual vehicle objects to which the vehicle number 10 is assigned at time t = T.

  Using the travel route information of the vehicle number 10 up to time t = T−1 indicated by the thick broken line in FIG. 4, among the output results of the vehicle detection unit 1, the vehicle object A indicated by the black circle “•” in FIG. 4. 4 is regarded as the observation position of the vehicle object of the vehicle number 10 at the time t = T (= the first route of the vehicle number 10), and the vehicle object indicated by the cross mark “×” in FIG. When the observation position of α is regarded as the observation position of the vehicle object of the vehicle number 10 at time t = T (= the second route of the vehicle number 10), the true vehicle position is estimated and the value is The result is output as a processing result of the travel route tracking unit 3.

とし、更に、任意の微小正数をεとし、また、時刻Ｔにおける車両実オブジェクトの前後関係に基づいて調整を行う係数をηとして、

と定義することにより、推定した予測位置と観測位置とのずれが小さいほど大きな値をとるスコア関数を採用することが可能である。 The route likelihood calculation unit 4 uses the true position of the vehicle object estimated by the travel route tracking unit 3 and the position of the observation data to calculate the probability probability and score of the estimated travel route as likelihood. . As the score S, for example, the predicted position and the observed position at the current time i of the vehicle real object estimated one hour ago are respectively

Further, an arbitrary minute positive number is ε, and a coefficient for adjustment based on the front-rear relationship of the vehicle real object at time T is η,

It is possible to adopt a score function that takes a larger value as the deviation between the estimated predicted position and the observed position is smaller.

  The vehicle number pruning unit 5 deletes the vehicle number corresponding to the route with the low likelihood calculated by the route likelihood calculating unit 4 from the vehicle numbers assigned to each vehicle object being tracked. In FIG. 2, the operation of the vehicle number pruning unit 5 in the actual vehicle object α will be described on the assumption that there is no vehicle that runs backward in the traveling direction. The travel route tracking unit 3 and the route likelihood calculation unit 4 use the travel route information of each vehicle number assigned to the vehicle real object α up to time t = T−1, and the observation position at time t = T is the vehicle. The true vehicle position and likelihood when it is regarded as the position of the actual vehicle object α obtained by the detection unit 1 is calculated for each vehicle number.

  Since the vehicle does not run backward, the likelihood of the vehicle numbers 10 and 21 inherited from the vehicle A traveling in front of the vehicle actual object α is reduced by reducing η in the equation (11). be able to. The vehicle number pruning unit 5 deletes the vehicle numbers 10 and 21 among the vehicle numbers assigned to the vehicle actual object α.

  In the vehicle detection unit 1 according to the present embodiment, a learned classifier such as a multilayer Neural Network, SVM (Support Vector Machine), or Adaboost can be used as a vehicle detection method. Alternatively, detection may be performed using an optical flow method, a background difference method, a motion stereo method, a compound eye stereo method, or the like.

  Further, in the vehicle number assignment unit 2 in the present embodiment, when the lane change is not considered, the inheritance of the vehicle number possessed by the actual vehicle object on the adjacent lane may be excluded. Further, the travel route tracking unit 3 in the present embodiment may use a statistical method for estimating a hidden state from observation data, such as HMM (Hidden Markov Model), Kalman Filter, DBN (Dynamic Bayesian Network). Good. Moreover, you may use the model which restrict | limits the behavior between adjacent vehicles like Cell Automaton.

  Furthermore, in the travel route tracking unit 3 according to the present embodiment, when there is a correspondence between the observation data and the vehicle object, a hidden state estimation method such as Kalman Filter, and when there is no correspondence, Cell A vehicle behavior restriction model such as Automaton may be used in combination. A travel route tracking method using Kalman Filter as a hidden state estimation method and Cell Automaton as a vehicle behavior restriction model will be described with reference to FIG.

  A vehicle object that is tracking a travel route based on observation data is referred to as a vehicle real object. In addition, a vehicle object that is disposed assuming that a vehicle exists between the actual vehicle objects is referred to as a vehicle virtual object. For a vehicle real object that is being tracked using the Kalman Filter, regardless of the presence of the vehicle virtual object, its position on the cell image based on the true vehicle position calculated from the highly likely travel route , And the vehicle virtual object is set as shown in FIG. 6 until the other vehicle real object existing behind the same lane.

  For the vehicle real object and the vehicle virtual object that are tracking the route using Cell Automaton, the position of the vehicle object on the cell image is set so that the vehicle real / virtual object existing in front of the same lane does not pass. Update.

  The likelihood used in the route likelihood calculation unit 4 in the present embodiment may be any function as long as it is a function that increases as the difference between the true position of the vehicle real object and the observation position decreases. .

  Details of the operation of the embodiment of the present invention will be described below with reference to the flowcharts of FIGS. FIG. 7 shows the overall processing flow. Step A1 in FIG. 7 shows the operation of the vehicle detection unit 1, steps A2 to A5 show the operation of the vehicle number assignment unit 2, and steps A7 to A8 show the travel route. The operation of the tracking unit 3 is shown, and steps A10 to A12 show the operation of the route likelihood calculating unit 4. Steps A6 and A9 represent the operation of the vehicle number pruning unit 5.

  FIG. 8 shows details of the processing flow in step A3 in FIG. 7, and FIG. 9 shows details of the operation in step B3 in FIG. 10 shows details of the operation of step B14 in FIG. 8, FIG. 11 shows details of the operation of step D4 in FIG. 10, and FIG. 12 shows details of the operation of step E9 in FIG.

  13 shows details of the operation of step E15 in FIG. 11, FIG. 14 shows details of the operation of step G5 in FIG. 13, and FIG. 15 shows details of the operation of step B15 in FIG. Further, FIG. 16 shows details of the operation of step A4 in FIG. 7, and FIG. 17 shows details of the operation of step A8 in FIG.

  Note that steps D2, D5, D10 in FIG. 10 and steps I2, I6 in FIG. 15 correspond to the operation of the vehicle number assignment unit 2, and step F4 in FIG. 12 and step H5 in FIG. It corresponds to the operation, and the others correspond to the operation of the travel route tracking unit 3.

  In the following description, when the true position (abbreviated as filtration position) one hour before the vehicle object corresponds to the observation position at the current time, or the duration of the missing measurement state is less than a predetermined threshold value Is Kalman Filter (hereinafter abbreviated as KM), otherwise Cell Automaton (hereinafter abbreviated as CA) is used, and the vehicle real object and vehicle virtual object shown in FIG. It is assumed that the behavior control of the vehicle object is performed.

  Also, FIGS. 18 and 19 show a structure for storing the travel route tracking information of each vehicle real object and each vehicle number, and a structure for storing information common to the processing is shown in FIG. Defined in In FIG. 18 and FIG. 19, the tracking state and the tracking mode represent the traveling route tracking state of the vehicle number of interest and the method used for the traveling route tracking. The tracking state is an initial state (= 0), a standby state (= 1), an execution state (= 2), and the tracking mode is an initial mode (= 0), a mode using KM (= 1), CA It is assumed that three modes (= 2) using, and transition between these states and modes as shown in FIG. MaxObj is the maximum number of vehicle numbers, and MaxPath is the maximum number of travel routes allowed for each vehicle number.

  In step A1, the observation data coordinates of the vehicle calculated in the coordinate system (image coordinates) on the image by the vehicle detection process are converted into a real scale coordinate system (world coordinates) set on the road, and at the current time. Stored in the structure od [1] [i] .info [j] .Obs. {X / y} as observation data coordinates. Note that i (= 0, 1, ..., LaneNum-1) is the lane number where the observation coordinates exist, and j (= 0, 1, ..., od [1] [i] .DatNum-1) is the actual vehicle object. Od [1] [i] .DatNum is the total number of observation coordinates existing on the lane of lane number i at the current time.

  In step A2, the observation coordinates od [1] [i] .info [j] .Obs. {X / y} of each vehicle at the current time (i = 0,1,..., LaneNum-1, j = 0,1 , ..., od [1] [i] .DatNum-1) and the predicted position od [0] [k] .info [l] .Prd. {X / y} (k = 0,1, ..., LaneNum-1, l = 0,1, ..., od [1] [k] .DatNum-1) are compared, and a pair whose value is other than a preset threshold is associated. When od [1] [i] .info [j] .Obs. {x / y} is associated with od [0] [k] .info [l] .Prd. {x, y} The correspondence is stored as od [1] [i] .info [j] .CorObsLane = k, od [1] [i] .info [j] .CorObsNo = l.

  Further, the value of the member of od [0] [k] .info [l] is copied to od [1] [i] .info [j]. In step A3, the actual vehicle object at the current time is tracked in step A4 for those objects that are not associated with the actual vehicle object at the current time among the actual vehicle objects one hour before.

  At step A5, paying attention to each vehicle real object, the assigned vehicle number, that is, od [i] [j] .info [k] .UseFlg [l] = 1 (i = 0,1 j = 0,... , LaneNum-1 k = 0,…, od [i] [j] .DatNum-1 l = 0,…, MaxObj-1) l path m (m = 0,…, MaxPath-1) Among the objects including the vehicle real object, the vehicle number ObjNo and the path number PathNo having the maximum score iif [l] .info [m] [1] .Score are calculated.

And
od [i] [j] .info [k] .CurPosi. {x / y} = iif [ObjNo] .info [PathNo] [1] .FltMean [0/1]
od [i] [j] .info [k] .Prd. {x / y} = iif [ObjNo] .info [PathNo] [1] .PrdMean [0/1]
od [i] [j] .info [k] .Obs. {x / y} = iif [ObjNo] .info [PathNo] [1] .Obs. {x / y}
od [i] [j] .info [k] .Velo. {x / y} = Average speed of the real object of interest from the start of tracking to the current time
od [i] [j] .info [k] .CurTrkStat = iif [ObjNo] .info [PathNo] [1] .TrkStat
od [i] [j] .info [k] .CurTrkMode = iif [ObjNo] .info [PathNo] [1] .TrkMode
od [i] [j] .info [k] .ObsLack = iif [ObjNo] .info [PathNo] [1] .ObsLack
Therefore, the tracking information based on the individual vehicle numbers is associated with the actual vehicle object.

  In step A6, a combination of a vehicle number and a travel route held by each vehicle real object is specified with a score value smaller than a predetermined threshold value, and the corresponding information is tracked for each vehicle number. In addition to deleting from the field iif, the value of the member array UseFlg of the tracking information structure od for each vehicle real object corresponding to the corresponding vehicle number is initialized to 0 as necessary.

  In step A7, the unrejected vehicle real objects one hour before and at the current time are stored again in the tracking information structure od [1] for each vehicle real object at the current time. In step A8, the setting of the vehicle virtual object and the update of the position coordinates on the cell images of all vehicle objects are performed, and the results are pc.CellImg [1], pc.ObjPosInfo [1], pc.ObjVeloInfo [1. ] Is saved.

  In step A9, the vehicle number and the travel route held by the actual vehicle object at the current time determined to be rejected are deleted from the corresponding tracking information structure for each vehicle number. In step A10, the actual vehicle object at the current time that was not treated as being rejected is stored in the tracking information structure od [0] for each actual vehicle object one hour before.

Regarding tracking status and tracking mode,
od [0] [i] .info [j] .PastTrkStat = od [1] [i] .info [j] .CurTrkStat
od [0] [i] .info [j] .PastTrkMode = od [1] [i] .info [j] .CurTrkMode
Update with.

  In step A11, the tracking information structure iif [i] .info [j] [1] (i = 0,..., MaxObj−1 j = 0,..., MaxPath−1) for each vehicle number at the current time is set to 1. Saved in the tracking information structure iif [i] .info [j] [0] for each vehicle number before the time. In step A12, cell information pc.CellImg [1], pc.ObjPosInfo [1], pc.ObjVeloInfo [1] at the current time are set as cell information one time before, pc.CellImg [0, pc.ObjPosInfo [0]. ], Pc.ObjVeloInfo [0], and initialize the cell information at the current time.

  The process of step A3 in FIG. 7 will be described with reference to FIG. Steps B8 and B11 are the same process. In step B1 in FIG. 8, the actual vehicle object structure at the current time is A = od [1] [i] .info [j], and steps B2 to B15 are performed for all possible i and j. . In step B2, the structure A is set as the input structure InCA to the CA, the vehicle actual object filtering position A.CurPosi at the current time is set to InCA.FltMean, and the velocity vector A.VeloCA is set to InCA.Velo.

  In step B3, the vehicle position NewPosi. {X / y} and the velocity vector NewVelo. {X / y} by CA are calculated for the input InCA. In step B4, the outputs NewPosi and NewVelo of step B3 are stored in A.PosCA and A.VeloCA, respectively. In step B5, TimeNo = 1, LaneNo = i, and ObsNo = j are set to identify the object of interest.

  In step B6, the tracking state TS and the tracking mode TM are set as TS = od [1] [i] .info [j] .CurTrkStat and TM = od [1] [i] .info [j] .CurTrkMode, respectively. . In step B7, a determination is made as to the state of correspondence between the vehicle real object of interest A and the vehicle real object one hour before, and the process moves to step B9 if a correspondence relationship exists, and to step B8 if no correspondence relationship exists. In step B8, the target vehicle real object A is changed to the tracking preparation state with A.CurTrkStat = 0 and A.CurTrkMode = 1. In step B9, the TM value is determined. If TM = 0, the process proceeds to step B10, and otherwise the process proceeds to B13.

  In step B10, the TS value is determined. If TS = 0, the process proceeds to step B11, and otherwise, the process proceeds to B12. In step B12, the target vehicle real object A is treated as a rejection by raising a flag. In step B13, the value of TM is determined. If TM = 1, the process proceeds to step B14, and otherwise, the process proceeds to B15. In step B14, with respect to all vehicle numbers (including the vehicle number newly generated at the current time) held by the target vehicle real object at the current time with A.Obs. {X / y} as the observation coordinates. , KM based travel route tracking. In step B15, tracking of the same travel route as in step B14 is performed using CA as a base.

  The operation in step B3 in FIG. 8 will be described with reference to FIG. In step C1, InCA.FltMean [0/1] representing the vehicle position is set to coordinates Pos1. {X / y}. In step C2, cell coordinates C1. {X / y} corresponding to Pos1. {X / y} are set on the cell image. In step C3, a cell image pc.CellImg [0] one hour before is searched ahead of the same x coordinate (that is, on the same lane) as C1, and pc.CellImg [0] [C2] ≠ 0 As described in FIG. 17, cell coordinates C2. {X / y} and a corresponding vehicle position Pos2. {X / y} corresponding to the cell in which the vehicle object exists are calculated.

  In step C4, a random number Random is generated, and in step C5, the y-direction velocity Velo.y = InCA.Velo.y + Random is set. In step C6, coordinates Pos3. {X / y} are set by the relational expression of Pos3.y = Pos1.y + Velo.y and Pos3.x = Pos1.x. In step C7, the positional relationship between Pos3 and Pos2 is determined. If Pos3 is ahead of Pos2, it means that it is impossible to overtake the vehicle on the same lane, so that Pos3.y = Pos2.y is set again in step C8. In step C9, Pos3. {X / y} is saved as NewPosi. {X / y}, NewVelo. {X / y} = Pos3. {X / y} -Pos1. {X / y}.

  Note that the addition of Random in Step C5 reproduces the acceleration / deceleration of the vehicle. Also, random numbers may be added for the x-direction speed, and the possibility of changing lanes may be taken into consideration in step C6.

  The process of step B14 in FIG. 8 will be described with reference to FIG. Note that steps D4, D8, and D12 are the same processing, steps D6 and D10 are the same processing, and steps D7 and D11 represent the same processing. In step D1, the TS value is determined. If TS = 0, the process proceeds to step D2, and otherwise the process proceeds to step D9.

  In step D2, an unused vehicle number at the current time is generated as a new vehicle number, and the value is set to NewNo. In step D3, the input vehicle number to KM is set to NewNo generated in step D2. In step D4, a travel route tracking process using KM or CA is executed on the input vehicle number using the observation coordinates A.Obs. {X / y}.

  In Step D5, in order to treat NewNo as an existing vehicle number, a flag is set as A.UseFlg [NewNo] = 1 and pc.UseFlg [NewNo] = 1, and a tracking state is set to an execution state as A.CurTrkStat = 2. Transition. In step D6, the vehicle actual object one hour before that exists before and after the self-propelled lane and the adjacent lane is searched based on the observation position A.Obs. {X / y} of the target vehicle real object A. And the vehicle number which the vehicle real object (refer description using FIG. 2) closest to the attention vehicle real object A in each lane and each direction (front and back) is inherited, and A for those vehicle numbers x Flag as .UseFlg [x] = 1 and consider it as an existing vehicle number.

  In step D7, the holding existing vehicle number of the actual vehicle object of interest A at the current time is set as the input number to KM. At step D9, the value of the missing state continuation frame number A.ObsLack of the real vehicle of interest at the current time is determined. If A.ObsLack ≠ 0, the process goes to step D10, otherwise the process goes to step D11. Each move.

  The process of step D4 in FIG. 10 will be described with reference to FIG. In step E1, a vehicle number i (= 0, 1,... MaxObj-1) is set, and the processes of steps E2 to E16 are executed for all i. In step E2, the value of pc.UseFlg [i] is determined. If the value is 1, the process proceeds to step E3 assuming that the vehicle number is being used for tracking the travel route. In Step E3, it is determined whether or not the vehicle number i is a newly generated vehicle number. If Yes, the search route upper limit number End is set to 1 in Step E4, and if No, End = in Step E5. Set to MaxPath.

  In step E6, a route number j (= 0, 1,..., End-1) is set, and steps E7 to E15 are performed for all j. In step E7, the tracking state one hour before the j-th route of the vehicle number i is TS = iif [i] .info [j] [0] .TrkStat, and the tracking mode is TM = iif [i] .info [ j] [0] .TrkMode. In step E8, the value of TM is determined. If the value is 1, the process goes to step E9. Otherwise, the process goes to step E10.

  In step E9, KM is executed using the jth travel route information iif [i] .info [j] [0] of the vehicle number i and given observation coordinates. In step E10, the TM value is determined. When the TM value is 2, the TS value is determined in steps E11 and E12. If TS = 1, mode = 1 is set in step E13, and if TS = 2, mode = 0 is set in step E14. In step E15, CA is executed using the jth travel route information iif [i] .info [j] [0] of the vehicle number i and the given mode value as the module execution mode.

  In step E16, the updated travel route tracking information iif [i] .info [j] [1] is rearranged in descending order of the score iif [i] .info [j] [1] .Score. As described later, the filtered position, predicted position, and score calculation result of the vehicle are stored in iif [i] .info [MaxPath] [1]. By performing step E16, iif [i] .info [j] [1] are always arranged in descending order of score, and it is easy to search for the maximum score route for each vehicle number performed in step A5 of FIG. become.

とする。ステップＦ２で、これらのベクトル、行列と与えられた観測座標を用いてろ過処理と予測処理とを行う。 The operation of step E9 in FIG. 11 will be described with reference to FIG. In step F1, filtration distribution average vector FltMean, variance-covariance matrix FltVar, prediction distribution average vector PrdMean, variance-covariance matrix of j-th travel route tracking information iif [i] .info [j] [0] of vehicle number i PrdVar is represented by the formulas (1) to (10), respectively.

And In step F2, filtration processing and prediction processing are performed using these vectors and matrices and given observation coordinates.

を、MaxPath+1 番目の走行経路追跡情報iif[i].info[MaxPath][1] のFltMean 、FltVar、PrdMean 、PrdVarにそれぞれ保存する。 In step F3, it was obtained in step F2.

Are stored in FltMean, FltVar, PrdMean and PrdVar of MaxPath + 1st travel route tracking information iif [i] .info [MaxPath] [1], respectively.

  In step F4, from iif [i] .info [j] [0] .PrdMean [0/1] (that is, the vehicle position at the current time predicted one time ago) and given observation coordinates, for example, the equation (11 ) Etc. are used to calculate the route score. In step F5, the score obtained in step F4 is stored in iif [i] .info [MatPath] [1] .Score, and tracking state iif [i] .info [MaxPath] [1] .TrkStat = 2 Mode iif [i] .info [MaxPath] [1] .TrkMode = 1 is set.

  If it is determined in step F6 that the target vehicle real object is associated with the vehicle real object one hour before, the process proceeds to step F7. If it is determined that the target vehicle real object is not compatible, the process proceeds to step F9. In step F7, the input observation coordinates are stored in iif [i] .info [MaxPath] [1] .Obs. {X / y}. In step F8, the value of iif [i] .info [MaxPath] [1] .ObsLack indicating the number of missing measurement state continuing frames is set to zero. In step F9, the value of iif [i] .info [j] [0] .ObsLack is copied to iif [i] .info [MaxPath] [1] .ObsLack, and then in step F10, iif [i]. Add 1 to info [MaxPath] [1] .ObsLack to update the number of missing frames.

  In step F11, since the target vehicle real object is in a missing state, that is, in a state where the corresponding observation coordinates do not exist, the 0th and 1st components of the average vector of the filtering process obtained in step F3 are regarded as the observation coordinates, Stored in iif [i] .info [MaxPath] [1] .Obs. {x / y}. If it is determined in step F12 that iif [i] .info [MaxPath] [1] .ObsLack matches the preset threshold value ObsLackThre, in order to shift the tracking mode to tracking by CA, in step F13, Set the value of tracking mode iif [i] .info [MaxPath] [1] .TrkMode to 2.

  The operation of step E15 in FIG. 11 will be described with reference to FIG. Note that step G7 is the same as step E9 in FIG. 11 and step G8 is the same as step E16 in FIG. In step G1, the predicted position iif [i] .info [j] [0] .PrdMean [0/1] obtained by the prediction process one time ago and the displacement vectors Δx and Δy between the input observation coordinates are respectively Calculated. In step G2, the value of the input mode is determined. If 0, the process proceeds to step G3, and otherwise, the process proceeds to step G4.

  In step G3, if both Δx and Δy are equal to or smaller than a preset threshold value, the process proceeds to step G6, and otherwise, the process proceeds to step G5. In step G4, the process moves to step G7 if both Δx and Δy are equal to or less than a preset threshold value, and to step G5 otherwise. In step G5, CA is executed on the target vehicle real object. In step G6, a process of changing the tracking state of the target vehicle real object being tracked by CA from the execution state to the standby state is performed.

  The operation in step G5 in FIG. 13 will be described with reference to FIG. Steps H7 and H8 in FIG. 14 are the same as steps F9 and F10 in FIG. In step H1, od [TimeNo] [LaneNo] .info [ObsNo] is a structure using TimeNo, LaneNo, ObsNo for identifying the actual vehicle object set in step B5 in FIG. 8 or step J5 in FIG. Copied to A. In step H2, the difference Disp. {X / y} between the given observation coordinates and the updated coordinates using CA is Disp. {X / y} = A.PosCA. {X / y} −A. Calculated as Obs. {X / y}.

  In step H3, the filtration processing distribution iif [i] .info [j] [0]. {FltMean, FltVar} one hour ago and the prediction processing distribution iif [i] .info [j] [0]. {PrdMean, PrdVar} is stored in iif [i] .info [MaxPath] [1]. {FltMean, FltVar} and iif [i] .info [MaxPath] [1]. {PrdMean, PrdVar}, respectively. In step H4, Disp.x is stored in iif [i] .info [MaxPath] [1]. {PrdMean, FltMean} [0/2], iif [i] .info [MaxPath] [1]. {PrdMean, Disp.y is added to FltMean} [1/3].

  In step H5, score calculation using, for example, Expression (11) is performed for the target travel route j of the target vehicle number i, and the value is stored in iif [i] .info [MaxPath] [1] .Score. . In step H6, the tracking state iif [i] .info [MaxPath] [1] .TrkStat at the current time is set to 2 indicating the tracking execution state, and the tracking mode iif [i] .info [MaxPath] [1] .TrkMode is set. Is set to 2 representing CA.

  In addition, since it is during the missing measurement state that the travel route is tracked by CA, the updated coordinates od [TimeNo] [LaneNo] .info [when the CA of the target vehicle real object is used as the input observation coordinates. ObsNo] .PosCA. {X / y} is used, and the value is stored in iif [i] .info [MaxPath] [1] .Obs. {X / y}. The reason why the process of step H4 is performed will be described. Since the filtration and prediction cannot be performed while the CA is in the execution mode, the vehicle object calculated by the CA in order to keep the positional relationship between the filtration position and the predicted position at the time of the tracking start by the CA temporally. Step H4 is a process of sequentially adding the positions of as offsets.

  The operation of Step G6 in FIG. 13 is the same as that in FIG. 14 except that the tracking state iif [i] .info [MaxPath] [1] .TrkStat is set to 1 representing the tracking standby state in Step H6 of FIG. Are the same.

  The operation of step B15 in FIG. 8 will be described with reference to FIG. Steps I2 to I8 and I9 to I10 in FIG. 15 are the same as steps D2 to D8 and D11 to D12 in FIG. In step I1, the value of TS is determined. If TS = 1, the process proceeds to step I2, and otherwise, the process proceeds to step I9.

  The operation of step A4 in FIG. 7 will be described with reference to FIG. Note that steps J2 to J4 in FIG. 16 are the same as steps B2 to B4 in FIG. 8, and steps J16 and J17 in FIG. 16 are the same as steps E9 and G5 in FIG. Is omitted. In step J1 of FIG. 16, steps J2 to J17 are performed for all possible i and j, with the vehicle real object structure one time ago as A = od [0] [i] .info [j]. Is called.

  In Step J5, TimeNo = 0, LaneNo = i, and ObsNo = j are set to identify the object of interest. In step J6, the tracking state TS and the tracking mode TM are set as TS = od [0] [i] .info [j] .CurTrkStat and TM = od [0] [i] .info [j] .CurTrkMode, respectively. . In step J7, a determination is made as to the state of correspondence between the vehicle real object of interest A and the vehicle real object one hour before. If no correspondence exists, the process moves to step J8. In step J8, the values of TS and TM are determined. If TM = 1 and TS = 0, the target vehicle real object is treated as a rejection in step J9, and otherwise the process proceeds to step J10.

  In step J10, a vehicle number k (= 0, 1,..., MaxObj-1) is set, and the processes of steps J11 to J17 are executed for all k. In step J11, the value of pc.UseFlg [k] is determined. If the value is 1, the processing is shifted to step J12 on the assumption that the vehicle number is being used for tracking the travel route. In Step J12, a path number l (= 0, 1,..., MaxPath-1) is set, and Steps J13 to J17 are performed for all l.

  In step J13, the tracking mode iif [k] .info [l] [0] .TrkMode of the l-th route of the vehicle number k one hour before is stored in tm. In step J14, if tm = 1, the process proceeds to step J16. Otherwise, the process proceeds to step J15. If it is determined in step J15 that tm = 2, the process proceeds to step J17.

  The operation of step A8 in FIG. 7 will be described with reference to FIG. Note that step K5 in FIG. 17 is the same as step B3 in FIG. In step K1, the vehicle position od [1] [i] .info [j] .CurPosi. {X / y} (i = 0, ..., LaneNum-1, j = 0, ..., od [ 1] [i] .DatNum-1) is obtained, and the value of the corresponding pixel pc.CellImg [1] [C] on the cell image at the current time is set to 1. In step K2, the processes of steps K3 to K7 are performed for all k (= 0,..., CellNum-1).

  In step K3, the pixel value pc.CellImg [0] [k] at the coordinate k in the cell image one time before is determined. If the value is 2, it is assumed that the vehicle virtual object exists on the corresponding cell coordinate. The process is moved to step K4. In step K4, the position information pc.ObjPosInfo [0] [k]. {X / y} and the speed information pc.ObjVeloInfo [0] [k]. {X / y} possessed by the vehicle virtual object on the cell coordinate k. Is set as the CA input InCA.

  In step K6, cell coordinates C ′ corresponding to the updated vehicle position coordinates NewPosi. {X / y} of the vehicle virtual object calculated in step K5 are calculated. In step K7, the updated vehicle virtual object information is pc.CellImg [1] [C '] = 2, pc.ObjPosInfo [1] [C']. {X / y} = NewPosi as current time information. . {x / y}, pc.ObjVeloInfo [1] [C ']. {x / y} = NewVelo. {x / y} In step K8, the processes of steps K9 to K13 are performed for all m (= 0,..., CellNum-1).

  In step K9, the value of the cell image pc.CellImg [1] [m] at the current time is determined. If the value is 1, the process moves to step K10. In step K10, the tracking mode at the current time of the vehicle real object existing on the cell coordinate m is stored in TrkMode. In step K11, a determination is made for the value of TrkMode. If the value is 1, the process moves to step K12. In step K12, a search is made for a cell in which the actual vehicle object exists behind the same x coordinate (that is, the same lane) as the cell coordinate m on the cell image pc.CellImg [1] at the current time, and the corresponding cell coordinate. Is stored as n.

  In step K13, pc.CellImg [1] [C ''] = 2, pc.ObjPosInfo [1] [C '' for all cell coordinates C '' sandwiched between cell coordinates m and cell coordinates n. ]. {x / y} = World coordinates corresponding to C ″, pc.ObjVeloInfo [1] [C ″]. {x / y} = Velocity vector of the actual vehicle object on cell coordinates m The vehicle virtual object is set (see FIG. 6).

  As described above, when it is difficult to uniquely assign a vehicle number such as a missing measurement state, the temporal continuity between the vehicle object and the vehicle number is interrupted by allowing the assignment of a plurality of vehicle numbers. It is possible to prevent this. In addition, it is robust to occlusion by controlling the likelihood calculation of the tracking route according to the presence or absence of the association between the vehicle object and the observation data, and pruning the assigned vehicle number based on the magnitude relationship of this likelihood. Highly accurate and highly efficient vehicle tracking can be performed.

  To summarize the above, according to the present invention, the vehicle continuity between the vehicle object and the vehicle number is prevented from being interrupted in the vehicle travel route tracking process, and the vehicle tracking that is robust to occlusion and highly accurate and highly efficient can be performed. It will be possible.

  It is obvious that each operation flow in the above embodiment can be configured such that the operation procedure is stored in advance in a recording medium as a program and is read and executed by a computer.

It is a general | schematic functional block diagram of embodiment of this invention. It is a figure explaining the multiple vehicle number allocation method to an attention vehicle object. It is a figure explaining the several path | route of the several vehicle number which an attention vehicle object holds. It is a figure explaining the true vehicle position estimated with respect to the several path | route of an attention vehicle number. It is a figure explaining the driving | running route tracking using a vehicle real object and a vehicle virtual object. It is a figure explaining the setting method of a vehicle virtual object. It is a flowchart which shows the whole operation | movement of embodiment of this invention. It is a flowchart which shows the detail of step A3 of FIG. It is a flowchart which shows the detail of step B3 of FIG. It is a flowchart which shows the detail of step B14 of FIG. It is a flowchart which shows the detail of step D4 of FIG. It is a flowchart which shows the detail of step E9 of FIG. It is a flowchart which shows the detail of step E15 of FIG. It is a flowchart which shows the detail of step G5 of FIG. It is a flowchart which shows the detail of step B15 of FIG. It is a flowchart which shows the detail of step A4 of FIG. It is a flowchart which shows the detail of step A8 of FIG. It is a graph explaining the tracking information structure for every vehicle real object. It is a chart explaining the tracking information structure for every vehicle number. It is a chart explaining a tracking process common information structure. It is a figure which shows the transition relationship of the tracking state with respect to an attention vehicle number, and tracking mode. It is a figure showing the concept of a cell image. It is a figure explaining the rule 184 of Cell Automaton.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Vehicle detection part 2 Vehicle number allocation part 3 Travel route tracking part 4 Path | route likelihood calculation part 5 Vehicle number pruning part

Claims (27)

Vehicle detection means for detecting individual vehicles as objects from road images;
Vehicle number assigning means for assigning a plurality of vehicle numbers to the detected individual objects;
A travel route tracking means for tracking a travel route based on each of the vehicle numbers assigned by the vehicle number assigning means;
Path likelihood calculating means for calculating the probability of each path as likelihood from the tracking path information obtained by the line path tracking means;
Vehicle number pruning means for deleting a vehicle number corresponding to a route having a smaller likelihood calculated by the route likelihood calculating means from among the vehicle numbers assigned by the vehicle number assigning means. Vehicle tracking device.   The route likelihood calculating means regards each of the assigned vehicle numbers as a temporally identified object having the vehicle number as the object of interest, The vehicle tracking device according to claim 1, wherein the likelihood is calculated as the likelihood based on a predicted position and a probability of a tracking route up to the current time.   The vehicle number pruning means excludes the vehicle number from the allocation target vehicle number when the likelihood corresponding to the allocated vehicle number falls below a preset threshold value. Or the vehicle tracking apparatus of 2.   4. The vehicle tracking device according to claim 2, wherein the traveling route tracking means switches the tracking method when vehicle observation data at the current time does not exist at the predicted position calculated one time ago.   Based on the positional relationship between the object of interest for which the observation data at the current time does not exist at the predicted position calculated one time ago and the other object with the correspondence between the predicted position and the observed data at the current time, The vehicle tracking device according to claim 4, wherein the filtering position of the object of interest at the current time is calculated.   The filtered position at the current time of the object of interest for which the observation data at the current time does not exist at the predicted position calculated one time before is the attention object among the other objects with the correspondence between the predicted position and the observation data at the current time. 6. The vehicle tracking device according to claim 5, wherein the vehicle tracking device is set so as not to overtake a filtering position of an object existing on the same lane as the object.   The vehicle tracking device according to claim 6, wherein the filtering position calculation of the target object based on the positional relationship between the target object and another object is performed according to an object movement model such as a cellular automaton. For a target object for which observation data at the current time does not exist at the predicted position calculated one time ago, when the duration of the unsupported state between the predicted position and the observed data is equal to or less than a preset threshold value, Considering this predicted position as observation data at the current time,
If it is larger than the threshold, another object with a correspondence between the predicted position and the observation data at the current time for the target object for which the observation data at the current time does not exist at the predicted position calculated one time ago The vehicle tracking device according to claim 4, wherein the filtering position at the current time of the object of interest is calculated based on the positional relationship between   4. The vehicle tracking device according to claim 2, wherein the tracking method is switched based on a difference between a predicted position calculated one hour before the object and a filtered position at the current time.   3. The vehicle tracking device according to claim 2, wherein a calculation method capable of predicting a position, such as a Kalman filter or a park filter, is used for calculating the filtering position at the current time and the predicted position one time ahead.   Multiple vehicle numbers are assigned only to objects that have transitioned from a state in which the predicted position of the object and the observation data do not correspond to a corresponding state, and the state in which the predicted position of the object corresponds to the observation data. 5. The vehicle tracking device according to claim 1, wherein for a continuing object, the vehicle number of the object at the previous time is taken over as the vehicle number of the object at the current time.   3. The vehicle tracking device according to claim 2, wherein the likelihood of the tracking route is calculated based on the filtering position calculated one time before.   7. The present invention introduces the existence of a virtual object between objects that are executing the travel route tracking, and sets so as not to overtake the filtering position of other objects existing on the same lane as the object of interest. The vehicle tracking device described. A vehicle detection step of detecting individual vehicles as objects from the road image;
A vehicle number assignment step for assigning a plurality of vehicle numbers to the detected individual objects;
A travel route tracking step for tracking the travel route based on each of the assigned vehicle numbers;
A path likelihood calculating step for calculating the probability of each path from the tracking path information as a likelihood; and
A vehicle tracking method comprising: a vehicle number pruning step for deleting a vehicle number corresponding to a route having a smaller likelihood calculated by the route likelihood calculating step among the assigned vehicle numbers.   The route likelihood calculating step regards each of the assigned vehicle numbers as an object in which the object of interest has the vehicle number and is identified with respect to time, and the filtering position at the current time, one hour ahead The vehicle tracking method according to claim 14, wherein the likelihood is calculated as the likelihood based on a predicted position and a probability of a tracking route up to the current time.   The vehicle number pruning step excludes the vehicle number from the allocation target vehicle number when the likelihood corresponding to the allocated vehicle number falls below a preset threshold value. Or the vehicle tracking method of 15.   The vehicle tracking method according to claim 15 or 16, wherein the traveling route tracking step switches the tracking method when vehicle observation data at the current time does not exist at the predicted position calculated one time ago.   Based on the positional relationship between the object of interest for which the observation data at the current time does not exist at the predicted position calculated one time ago and the other object with the correspondence between the predicted position and the observed data at the current time, The vehicle tracking method according to claim 17, wherein the filtering position of the object of interest at the current time is calculated.   The filtered position at the current time of the object of interest for which the observation data at the current time does not exist at the predicted position calculated one time before is the attention object among the other objects with the correspondence between the predicted position and the observation data at the current time. 19. The vehicle tracking method according to claim 18, wherein the vehicle tracking method is set so as not to overtake a filtering position of an object existing on the same lane as the object.   The vehicle tracking method according to claim 19, wherein the filtering position calculation of the target object based on a positional relationship between the target object and another object is performed according to an object movement model such as a cellular automaton.   For a target object for which observation data at the current time does not exist at the predicted position calculated one time ago, when the duration of the unsupported state between the predicted position and the observed data is equal to or less than a preset threshold value, The vehicle tracking method according to claim 17, wherein the predicted position is regarded as observation data at the current time, and if the predicted position is larger than a threshold value, the method according to claim 18 is used.   The vehicle tracking method according to claim 15 or 16, wherein the tracking method is switched based on a difference between a predicted position calculated one hour before the object and a filtered position at the current time.   16. The vehicle tracking method according to claim 15, wherein a calculation method capable of predicting a position, such as a Kalman filter or a park filter, is used for calculating the filtering position at the current time and the predicted position one time ahead.   Multiple vehicle numbers are assigned only to objects that have transitioned from a state in which the predicted position of the object and the observation data do not correspond to a corresponding state, and the state in which the predicted position of the object corresponds to the observation data. 18. The vehicle tracking method according to claim 14, wherein for a continuing object, the vehicle number of the object at the previous time is taken over as the vehicle number of the object at the current time.   16. The vehicle tracking method according to claim 15, wherein the likelihood of the tracking route is calculated based on the filtering position calculated one time before.   The virtual object existence is introduced between the objects that are executing the travel route tracking, and the filtering position of other objects existing on the same lane as the object of interest is set not to be overtaken. The vehicle tracking method described. A program for executing a vehicle tracking method by a computer,
Processing for detecting individual vehicles as objects from road images;
A process of assigning a plurality of vehicle numbers to the detected individual objects;
A process of tracking the travel route based on each of the assigned vehicle numbers;
A process of calculating the probability of each route from the tracking route information as a likelihood,
A computer-readable program comprising: deleting a vehicle number corresponding to a route having a small likelihood calculated by the route likelihood calculating step from among the assigned vehicle numbers.

Images