Complex multi-target adaptive detection method for smart transportation
By constructing a three-dimensional scene-oriented object detection model in the vehicle pedestrian detection algorithm and using a threshold adaptive adjustment mechanism, combining time and weather information for model selection and fusion, the problem of detection accuracy in complex traffic environments in the existing technology is solved. It has achieved improvement in the accuracy of multi-objective detection throughout the day and around the clock.
Patent Information
- Application Number
- CN202210336720.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-03-31
AI Technical Summary
The existing vehicle and pedestrian detection algorithms are difficult to achieve accurate detection all-day and all-weather in complex traffic environments, and ignore scene information, resulting in frequent occurrence of false detection and missed detection.
The YOLOv3 model is used to build a three-dimensional scene-oriented object detection model, and the corresponding model is selected based on the current time and weather information, and the multi-dimensional model is fusion through the threshold adaptive adjustment mechanism to output the final category and its probability.
It has achieved the accuracy of all-day and all-weather multi-objective detection in complex traffic environments, reducing false detection and missed detection, and improving detection accuracy.
Smart Images

Figure CN114842427B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a multi-target adaptive detection method, and in particular to a complex multi-target adaptive detection method for smart transportation. Background Art
[0002] With the rapid development of computer vision technology, machine vision detection technology based on neural networks has made breakthrough progress, among which vehicle and pedestrian detection algorithms based on deep learning have gradually entered the field of vision of many researchers. Vehicle and pedestrian detection algorithms are described as the problem of automatically identifying and locating vehicles and pedestrians in images or videos.
[0003] In the problem of image target behavior recognition, the scene is an important element in the image. Based on the scene attributes, we can judge the behaviors that are unlikely to occur or impossible in the image. For example, people often do not wash clothes in the bedroom, and pedestrians often do not appear on the motor vehicle lane. These knowledge are all prior knowledge. However, the existing vehicle and pedestrian detection algorithms often ignore this scene information, which poses a severe challenge to the accurate detection of various targets all day and all weather in complex traffic environments. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to provide a complex multi-target adaptive detection method for smart transportation that can realize all-day and all-weather detection of various targets.
[0005] Technical solution: The complex multi-target adaptive detection method of the present invention comprises the following steps:
[0006] S1, the basic model uses YOLOv3, which is trained with given time class, weather class, and target class to obtain a target detection model for three-dimensional scenes;
[0007] S2, according to the current time type and weather type, select the corresponding target detection model for the three-dimensional scene; use the target detection model for the three-dimensional scene to detect the target object in the picture, determine the number of target detection models, and complete the selection of the multi-dimensional model;
[0008] The threshold adaptive adjustment mechanism is used as the standard for the fusion of multi-dimensional model detection results;
[0009] S3, based on the threshold adaptive adjustment mechanism, performs multi-dimensional model fusion and outputs the final category of the target to be detected and its corresponding probability.
[0010] Furthermore, in step S1, the specific implementation process of constructing a target detection model for a three-dimensional scene is as follows:
[0011] S101, build a three-dimensional scene space
[0012] The three-dimensional scenes are: time category, weather category, and target category, wherein the time category is divided into early morning / evening / ...m1 category, the weather category is divided into sunny / cloudy / ...m2 category, and the target category is divided into motor vehicles, non-motor vehicles, pedestrians, others...m3+1 category, where motor vehicles, non-motor vehicles, pedestrians... have a total of m3 categories, and the other 1 category refers only to others;
[0013] S102, select sample data
[0014] Given the time class, weather class, and target class, select the corresponding sample data;
[0015] S103, sample annotation
[0016] Use the data labeling tool labelimg to label the collected sample data;
[0017] S104, model training
[0018] The labeled data set is trained with the YOLOv3 model to obtain m1×m2×m3 target detection models for three-dimensional scenes.
[0019] Furthermore, in step S2, the selection of the multi-dimensional model is implemented as follows:
[0020] Taking the scenes of the two dimensions of time and weather as prior knowledge, when it is necessary to detect the target object in the currently taken picture, the central server in the intelligent transportation system is first used to call the calendar and meteorological monitoring points to obtain the current time and weather conditions, and then the corresponding trained target detection model for three-dimensional scenes is selected according to the current time class and weather class. Finally, the obtained target detection model for three-dimensional scenes is used to detect the target object in the picture; the number of selected models is m3.
[0021] Furthermore, in step S2, the implementation process of the threshold adaptive adjustment mechanism is as follows:
[0022] Define two thresholds: false positive threshold d1 and intersection-over-union threshold U th , the calculation formula of the false detection threshold d1 is as follows:
[0023] d1=F(BEV)=d+δ×BEV
[0024] Among them, d is the false detection basic threshold, δ is the coefficient, and BEV is the baseline environment change;
[0025] Intersection-over-union ratio threshold U th The calculation formula is as follows:
[0026] U th =H(BEV)=U-ε×BEV
[0027] Among them, U is the basic threshold of intersection-over-union ratio, and ε is the coefficient;
[0028] Define the baseline environment change:
[0029]
[0030] Among them, N represents the number, Current represents the current scene, which refers to the image to be detected; Base represents the reference scene, which refers to the image of a sunny day; R, G, B represent the three components of the color space of an image, red, green, and blue, and the value range is [0,255], which is an integer; i represents the value of the three components of R, G, and B. Respectively represent the number of i values of the three components R, G, and B in the current scene, Respectively represent the number of i for the three components of the reference scene R, G, and B, and N CurrentRGB +N BaseRGB Indicates the total number of R, G, and B components in the current scene and the reference scene.
[0031] Furthermore, in step S3, the implementation process of multi-dimensional model fusion is as follows:
[0032] S3.1, let the number of models in the target detection model for three-dimensional scenes selected in step S2 be 3. These three models detect the current sample respectively, and mark the prediction box, corresponding category and probability on the target to be detected; the category result outputs for the target to be detected are j1 / other, j2 / other, j3 / other, respectively, where j1 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles and pedestrians in the target class, j2 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles and pedestrians in the target class, and j3 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles and pedestrians in the target class, but j1, j2 and j3 are different from each other. The probabilities of the target objects selected by the prediction box belonging to j1, j2 and j3 are respectively
[0033] S3.2, the prediction boxes of the three model prediction target outputs in step S3.1 are fused and the output results are superimposed on a picture, which is divided into the following three cases:
[0034] A. If a prediction box e has no overlapping prediction box, determine the category probability on the prediction box e k is an integer, and its value is different from the false detection threshold d1:
[0035] like Then the prediction box has no false detection, and the output is still the corresponding category and probability marked by it
[0036] like Then the prediction box is misdetected, and then the prediction box is deleted;
[0037] B. If there are two prediction boxes e and f overlapping each other, calculate the intersection and union ratio U of the two prediction boxes ef and the intersection-combination ratio threshold U th For comparison, there are two situations:
[0038] B11. If U ef ≥U th
[0039] Then merge the two prediction boxes e and f into one prediction box. If the corresponding category results of the prediction boxes e and f are others, others, others, j2, j1, and others respectively, then the output categories of the fused prediction boxes are others, j2, and j1 respectively.
[0040] If the corresponding category results on the prediction box e and the prediction box f are j1 and j2, the correction probabilities of category j1 and category j2 are compared. When the two prediction boxes are fused, the correction probabilities of j1 and j2 are defined as follows:
[0041]
[0042]
[0043] in, are the numbers of targets j1 and j2 output by the corresponding models respectively;
[0044] like The output target category of the fused prediction box is determined to be j1, otherwise it is j2;
[0045] B12. If U ef th
[0046] Then the two prediction boxes e and f are not merged, and the two prediction boxes are processed separately as if there is no prediction box overlapping with them, and the processing is performed according to step A;
[0047] C. If three prediction boxes e, f, and g overlap with each other, calculate the intersection-and-union ratio U between the prediction boxes ef , U eg , U fg , which can be divided into two situations:
[0048] C11. If there are two or more intersection ratios greater than U th
[0049] Then merge these three prediction boxes into one prediction box. If the corresponding category results of prediction box e, prediction box f, and prediction box g are others, others, others, others, j3, j1, others, others, others, j2, and others, respectively, then the output categories of the fused prediction box are others, j3, j1, and j2 respectively.
[0050] If the corresponding category results on the prediction box e, prediction box f, and prediction box g are j1, j2, others, j1, others, j3, and others, j2, j3, respectively, then compare the corrected probabilities of the two corresponding target categories. and and and Then the output category of the fused prediction box is the category with a larger correction probability among the two target categories;
[0051] If the corresponding categories on the prediction boxes e, f, and g are j1, j2, and j3, then compare the corrected probabilities of the three target categories. Then the output category of the fused prediction box is the category with the larger correction probability among the three target categories; among them, is the number of targets j3 output by the corresponding model;
[0052] C12. If there is only one intersection-combination ratio greater than U th
[0053] The corresponding two prediction boxes are merged into one prediction box, and the fusion process of the two prediction boxes is processed according to step B, and the other prediction box is processed according to step A.
[0054] Compared with the prior art, the present invention has the following significant effects:
[0055] 1. The present invention selects the corresponding target detection model for three-dimensional scenes according to the current time class and weather class, and proposes a threshold adaptive adjustment mechanism to adaptively fuse the detection results of multiple models and output the final category of the target to be detected and its corresponding probability.
[0056] 2. The present invention can be applied to the multi-target adaptive precision detection driven by a priori in the field of machine vision in all-day and all-weather environments, and has a very broad application prospect. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 It is the overall structural diagram of the present invention;
[0058] Figure 2 A structural block diagram of the object detection model for three-dimensional scenes in the present invention;
[0059] Figure 3 It is a structural block diagram of the prior-driven multi-dimensional model adaptive joint detection in the present invention;
[0060] Figure 4(a) shows the detection result of the YOLOv3 model for the image to be detected.
[0061] Figure 4(b) Detection results of the multi-target adaptive detection model for the image to be detected. DETAILED DESCRIPTION
[0062] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0063] The present invention applies scene information to vehicle and pedestrian detection algorithms to achieve accurate detection of various types of targets around the clock and in all weather conditions in complex traffic environments.
[0064] like Figure 1 As shown, the specific process of the complex multi-target adaptive detection method of the present invention is as follows:
[0065] Step S1, constructing a target detection model for a three-dimensional scene;
[0066] The basic model uses YOLOv3, which is trained with given time, weather, and target classes to obtain a target detection model for three-dimensional scenes.
[0067] Step S2, selection of multidimensional model;
[0068] According to the current time and weather categories, a corresponding target detection model for three-dimensional scenes is selected for target detection, i.e., selection of a multi-dimensional model;
[0069] A threshold adaptive adjustment mechanism is proposed as the criterion for fusing the detection results of m3 models;
[0070] Step S3, multi-dimensional model fusion;
[0071] Based on the threshold adaptive adjustment mechanism, multi-target adaptive joint detection, i.e. multi-dimensional model fusion, is performed to output the final category of the target to be detected and its corresponding probability.
[0072] like Figure 2 As shown, in step S1, a target detection model for a three-dimensional scene is constructed, as follows:
[0073] Step S101, constructing a three-dimensional scene space;
[0074] Aiming at the complex traffic environment: different time scenes such as day / night / early morning / evening, etc., different weather scenes such as sunny / cloudy / rainy / snowy, etc., and multiple traffic targets: motor vehicles, non-motor vehicles, pedestrians and other target types, a three-dimensional scene space thinking method is proposed. The three dimensions of the scenes are: time, weather, and target type. The time category can be divided into early morning / evening / …m1 category, the weather category can be divided into sunny / cloudy / …m2 category, and the target category can be divided into motor vehicles, non-motor vehicles, pedestrians, others…m3+1 category (where 1 category refers to others). Here, there are a total of m3 categories of motor vehicles, non-motor vehicles, pedestrians... and the other 1 category only refers to others.
[0075] Step S102, selecting sample data;
[0076] Given a time class, a weather class, and a target class, select the corresponding sample data, where the target class does not include the other category (because the target class is divided into m3+1 categories, where the single 1 category refers to others, and others may be pets, plastic bags, etc., which are not vehicle or pedestrian target recognition objects, but do exist in actual scenarios; they are used to distinguish from the recognition target, or as a reference for target recognition, so when training the model, only the m3 category targets are targeted, and the other category is not included).
[0077] Step S103, sample labeling;
[0078] The data labeling tool labelimg is used to label the collected sample data.
[0079] Step S104, model training;
[0080] The labeled data set is trained with the YOLOv3 model, and a total of m1×m2×m3 target detection models for three-dimensional scenes can be obtained.
[0081] like Figure 3 The flowchart of the prior-driven multi-dimensional model adaptive joint detection method is shown, which mainly includes the following steps:
[0082] Step S201, selection of a multi-dimensional model;
[0083] Taking the scenes of the two dimensions of time and weather as prior knowledge, when it is necessary to detect the target object in the currently taken picture, the central server in the intelligent transportation system can first call the calendar and meteorological monitoring points to obtain the current time and weather conditions, and then select the corresponding trained target detection model for three-dimensional scenes according to the current time and weather types.
[0084] Step S202, several multi-dimensional models perform target detection;
[0085] The target object in the image is detected using the target detection model for the three-dimensional scene obtained in step S201, wherein the number of selected models is m3.
[0086] Step S203, threshold adaptive adjustment mechanism;
[0087] Define two thresholds: false positive threshold d1 and intersection-over-union threshold U th , these two thresholds determine the accuracy of joint detection, including false detection rate and missed detection rate. The calculation formula of false detection threshold d1 is as follows:
[0088] d1=F(BEV)=d+δ×BEV (1)
[0089] Among them, d is the false detection basic threshold, δ is the coefficient, BEV is the baseline environment change,
[0090] Intersection-over-union ratio threshold U th The calculation formula is as follows:
[0091] U th =H(BEV)=U-ε×BEV (2)
[0092] Among them, U is the basic threshold of intersection-over-union ratio, ε is the coefficient,
[0093] Generally speaking, the image of a sunny day scene is used as the reference scene. The detection in the reference scene is relatively accurate. If the current scene changes more compared with the reference scene, the false detection degree will be higher. The accuracy of the same target detection will decrease, and the corresponding prediction box offset will increase. Therefore, the greater the change in the reference environment, the larger d1 should be to reduce false detection and U th The smaller it is, the lower the deviation of the same target is, and the change in the baseline environment is defined as:
[0094]
[0095] Where N represents the number, Current represents the current scene, which refers to the image to be detected; Base represents the reference scene, which refers to the image of a sunny day; R, G, B represent the three components of the color space of an image, red, green, and blue, with a value range of [0,255], taking integers, and i represents the values of the three components of R, G, and B. Respectively represent the number of i values of the three components R, G, and B in the current scene, Respectively represent the number of i for the three components of the reference scene R, G, and B, and N CurrentRGB +N BaseRGB Indicates the total number of R, G, and B components in the current scene and the reference scene, used for normalization.
[0096] Step S204, multi-dimensional model fusion;
[0097] The multi-dimensional fusion process is as follows:
[0098] (1) On the one hand, in order to briefly describe the fusion process of the prediction frame, on the other hand, since the targets detected in the real environment can be mainly divided into three categories: motor vehicles, non-motor vehicles, and pedestrians, let m3=3 in the m3 target detection models for three-dimensional scenes selected in step S2. These three models detect the current sample respectively, mark the prediction frame, the corresponding category and the probability on the target to be detected, and the category result outputs for the target to be detected are j1 / other, j2 / other, and j3 / other, respectively, where j1 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles, and pedestrians in the target class, j2 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles, and pedestrians in the target class, and j3 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles, and pedestrians in the target class, but j1, j2, and j3 are different from each other. The probabilities of the target objects selected by the prediction frame belonging to j1, j2, and j3 are respectively
[0099] (2) The prediction boxes of the target outputs of the three models in (1) are fused and the output results are superimposed on an image. There are three cases:
[0100] A. If a prediction box e has no overlapping prediction box, determine the category probability on the prediction box e k is an integer, and its value is greater than the false detection threshold d1. Then the prediction box has no false detection, and the output is still the corresponding category and probability marked by it like Then the prediction box is misdetected, and then the prediction box is deleted;
[0101] B. If there are two prediction boxes e and f overlapping each other, calculate the intersection and union ratio U of the two prediction boxes ef and the intersection-combination ratio threshold U th For comparison, there are two situations:
[0102] B11. If U ef ≥U th
[0103] Then merge the two prediction boxes e and f into one prediction box. If the corresponding category results of the prediction boxes e and f are others, others, others, j2, j1, and others, respectively, the output categories of the fused prediction boxes are others, j2, and j1. If the corresponding category results of the prediction boxes e and f are j1 and j2, compare the corrected probabilities of category j1 and category j2. When the two prediction boxes are fused, the corrected probabilities of j1 and j2 are defined as follows:
[0104]
[0105]
[0106] in, are the number of targets j1 and j2 output by the corresponding models. If the number of targets of a certain type is large, it means that the scene is a scene with a large number of targets of this type, and its probability increases accordingly (also driven by a priori, for example, the probability of a car in the fast lane is higher).
[0107] like The output target category of the fused prediction box is determined to be j1, otherwise it is j2;
[0108] B12. If U ef th
[0109] The two prediction boxes are not merged, and the two prediction boxes are processed separately as if there is no prediction box overlapping with them, that is, they are processed according to step A;
[0110] C. If three prediction boxes e, f, and g overlap with each other, calculate the intersection-and-union ratio U between the prediction boxes ef , U eg , U fg , which can be divided into two situations:
[0111] C11. If there are two or more intersection ratios greater than U th
[0112] Then merge these three prediction boxes into one prediction box. If the corresponding category results of prediction box e, prediction box f, and prediction box g are others, others, others, others, j3, j1, others, others, j2, and others, respectively, then the output categories of the fused prediction box are others, j3, j1, and j2. If the corresponding category results of prediction box e, prediction box f, and prediction box g are j1, j2, others, j1, others, j3, and others, j2, and j3, respectively, then compare the corrected probabilities of the two corresponding target categories. and and and Then the output category of the fused prediction box is the category with the larger correction probability among the two target categories. If the corresponding categories on the prediction boxes e, f, and g are j1, j2, and j3, the correction probabilities of the three target categories are compared. Then the output category of the fused prediction box is the category with the larger correction probability among the three target categories; among them, is the number of targets j3 output by the corresponding model;
[0113] C12. If there is only one intersection-combination ratio greater than U th
[0114] The corresponding two prediction boxes are merged into one prediction box, and the fusion process of the two prediction boxes is processed according to step B, and there is a single prediction box processed according to step A.
[0115] In the experimental verification of the present invention, the targets to be detected are divided into two categories: cars and non-motor vehicles (nmv). All targets other than the targets to be detected are marked as other. As shown in Figure 4 (a), it is the detection result diagram of the YOLOv3 model for the image to be detected. It can be seen that when there is no three-dimensional scene space constructed for time, weather, and target type, the detection result of the YOLOv3 model can identify the vehicles with overlapping parts separately, but there is also a phenomenon of missed detection. As shown in Figure 4 (b), it is the detection result diagram of the multi-target adaptive detection model for the image to be detected. It can be seen that the multi-target adaptive detection model can identify the vehicles with overlapping parts separately without any false detection or missed detection. Table 1 shows the evaluation indicators of the two models. It can be seen that compared with the precision and recall rate in the YOLOv3 model, the precision and recall rate in the multi-target adaptive joint detection model are increased by 3.46% and 2.68% respectively. The mAP value of the YOLOv3 network model is 74.37%. When the adaptive coefficient δ is 1 and ε is 1, the threshold adaptive adjustment mechanism is introduced. At this time, the mAP value of the model is 77.86%. Compared with the mAP value of the YOLOv3 model, the mAP value of the multi-target adaptive joint detection model is increased by 1.35%.
[0116] Table 1 Evaluation indicators of the two models
[0117]
[0118] From this, we can conclude that after the introduction of the threshold adaptive adjustment mechanism, the mAP value of the multi-target adaptive joint detection model has been improved, that is, the model performance has been improved, and it can achieve accurate detection of multiple targets all day and all weather in complex traffic scenarios.
Claims
1. A complex multi-target adaptive detection method for smart transportation, characterized in that: The steps include: S1, the basic model uses YOLOv3, which is trained with given time class, weather class, and target class to obtain a target detection model for three-dimensional scenes; S2, according to the current time type and weather type, select the corresponding target detection model for the three-dimensional scene; use the target detection model for the three-dimensional scene to detect the target object in the picture, determine the number of target detection models, and complete the selection of the multi-dimensional model; The threshold adaptive adjustment mechanism is used as the standard for the fusion of multi-dimensional model detection results; S3, based on the threshold adaptive adjustment mechanism, multi-dimensional model fusion is performed to output the final category of the target to be detected and its corresponding probability; the implementation process of the multi-dimensional model fusion is as follows: S3.1, let the number of models in the target detection model for three-dimensional scenes selected in step S2 be 3. These three models detect the current sample respectively, and mark the prediction box, corresponding category and probability on the target to be detected; the category result outputs for the target to be detected are j1 / other, j2 / other, j3 / other, respectively, where j1 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles and pedestrians in the target class, j2 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles and pedestrians in the target class, and j3 indicates that the category of the target to be detected is any one of the motor vehicles, non-motor vehicles and pedestrians in the target class, but j1, j2 and j3 are different from each other. The probabilities of the target objects selected by the prediction box belonging to j1, j2 and j3 are respectively S3.2, the prediction boxes of the three model prediction target outputs in step S3.1 are fused and the output results are superimposed on a picture, which is divided into the following three cases: A. If a prediction box e has no overlapping prediction box, determine the category probability on the prediction box e k is an integer, and its value is different from the false detection threshold d1: like Then the prediction box has no false detection, and the output is still the corresponding category and probability marked by it like Then the prediction box is misdetected, and then the prediction box is deleted; B. If there are two prediction boxes e and f overlapping each other, calculate the intersection and union ratio U of the two prediction boxes ef and the intersection-combination ratio threshold U th For comparison, there are two situations: B11. If U ef ≥U th Then merge the two prediction boxes e and f into one prediction box. If the corresponding category results of the prediction boxes e and f are others, others, others, j2, j1, and others respectively, then the output categories of the fused prediction boxes are others, j2, and j1 respectively. If the corresponding category results on the prediction box e and the prediction box f are j1 and j2, the correction probabilities of category j1 and category j2 are compared. When the two prediction boxes are fused, the correction probabilities of j1 and j2 are defined as follows: in, are the numbers of targets j1 and j2 output by the corresponding models respectively; like The output target category of the fused prediction box is determined to be j1, otherwise it is j2; B12. If U ef th Then the two prediction boxes e and f are not merged, and the two prediction boxes are processed separately as if there is no prediction box overlapping with them, and the processing is performed according to step A; C. If three prediction boxes e, f, and g overlap with each other, calculate the intersection-and-union ratio U between the prediction boxes ef , U eg , U fg , which can be divided into two situations: C11. If there are two or more intersection ratios greater than U th Then merge these three prediction boxes into one prediction box. If the corresponding category results of prediction box e, prediction box f, and prediction box g are others, others, others, others, j3, j1, others, others, others, j2, and others, respectively, then the output categories of the fused prediction box are others, j3, j1, and j2 respectively. If the corresponding category results on the prediction box e, prediction box f, and prediction box g are j1, j2, others, j1, others, j3, and others, j2, j3, respectively, then compare the corrected probabilities of the two corresponding target categories. and and and Then the output category of the fused prediction box is the category with a larger correction probability among the two target categories; If the corresponding categories on the prediction boxes e, f, and g are j1, j2, and j3, then compare the corrected probabilities of the three target categories. Then the output category of the fused prediction box is the category with the larger correction probability among the three target categories; among them, is the number of targets j3 output by the corresponding model; C12. If there is only one intersection-combination ratio greater than U th The corresponding two prediction boxes are merged into one prediction box, and the fusion process of the two prediction boxes is processed according to step B, and the other prediction box is processed according to step A.
2. The complex multi-target adaptive detection method for smart transportation according to claim 1 is characterized in that: In step S1, the specific implementation process of constructing a target detection model for a three-dimensional scene is as follows: S101, constructing a three-dimensional scene space The three-dimensional scenes are: time category, weather category, and target category, wherein the time category is divided into early morning / evening / ...m1 category, the weather category is divided into sunny / cloudy / ...m2 category, and the target category is divided into motor vehicles, non-motor vehicles, pedestrians, others...m3+1 category, where motor vehicles, non-motor vehicles, pedestrians... have a total of m3 categories, and the other 1 category refers only to others; S102, select sample data Given the time class, weather class, and target class, select the corresponding sample data; S103, sample annotation Use the data labeling tool labelimg to label the collected sample data; S104, model training The labeled data set is trained with the YOLOv3 model to obtain m1×m2×m3 target detection models for three-dimensional scenes.
3. The complex multi-target adaptive detection method for smart transportation according to claim 1 is characterized in that: In step S2, the selection process of the multi-dimensional model is as follows: Taking the scenes of the two dimensions of time and weather as prior knowledge, when it is necessary to detect the target object in the currently captured picture, the central server in the intelligent transportation system first calls the calendar and meteorological monitoring points to obtain the current time and weather conditions, and then selects the corresponding trained target detection model for three-dimensional scenes according to the current time class and weather class. Finally, the obtained target detection model for three-dimensional scenes is used to detect the target object in the picture; The number of models selected is m3.
4. The complex multi-target adaptive detection method for smart transportation according to claim 1 is characterized in that: In step S2, the implementation process of the threshold adaptive adjustment mechanism is as follows: Define two thresholds: false positive threshold d1 and intersection-over-union threshold U th , the calculation formula of the false detection threshold d1 is as follows: d1=F(BEV)=d+δ×BEV Among them, d is the false detection basic threshold, δ is the coefficient, and BEV is the baseline environment change; Intersection-over-union ratio threshold U th The calculation formula is as follows: IN th =H(BEV)=U-ε×BEV Among them, U is the basic threshold of intersection-over-union ratio, and ε is the coefficient; Define the baseline environment change: Among them, N represents the number, Current represents the current scene, which refers to the image to be detected; Base represents the reference scene, which refers to the image of a sunny day; R, G, B represent the three components of the color space of an image, red, green, and blue, and the value range is [0,255], which is an integer; i represents the value of the three components of R, G, and B. Respectively represent the number of i values of the three components R, G, and B in the current scene, Respectively represent the number of i for the three components of the reference scene R, G, and B, and N CurrentRGB +N BaseRGB Indicates the total number of R, G, and B components in the current scene and the reference scene.
