Video-based lodging traffic cone automatic righting system

Abstract

The invention discloses a video-based lodging traffic cone automatic righting system, which comprises an identification module, an information processing module and a righting module, the identification module collects traffic cone videos through a monocular camera, the information processing module makes a data set containing different traffic cone states, and an optimal pre-training model is obtained by using an improved YOLOv8 network; and according to the information output by the information processing module, the righting module is controlled to perform rotary righting and righting operation on the traffic cone. The state and angle of the traffic cone are recognized from a righting mode, a rotary gathering and righting function and a righting function are integrated, the requirement of traffic cone righting for lodging angle recognition precision is lowered through the design of a righting mechanism, and the problem that the righting mechanism interferes with continuous recovery of the traffic cone after the traffic cone is righting is solved; the problem that temporary parking is needed when the traffic cones fall down in the righting state before recovery is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of road maintenance equipment, and in particular relates to a video-based automatic righting system for fallen traffic cones. Background Art

[0002] With the annual increase in expressway mileage and the rapid aging of road infrastructure, the number of routine maintenance, preventive and repair maintenance projects, and the construction, maintenance, and repair of road-related electromechanical equipment on my country's expressways will continue to increase in the future. In this context, traffic cones, as critical infrastructure for expressway operations, are subject to significant risks, numerous casualties, and slow operation speeds when manually deployed. Semi-automatic and fully automatic deployment and retraction systems can significantly alleviate the challenges encountered during manual operations and improve worker safety. Considering that traffic cones are susceptible to falling over in the control area due to the influence of oncoming vehicles and weather factors, this hinders the automated recovery of traffic cones, renders them ineffective as warning devices, and interferes with passing vehicles, even posing a serious threat to traffic safety. Therefore, implementing an automatic righting function for fallen traffic cones is of great significance.

[0003] The existing righting system for fallen traffic cones has certain problems due to its own structure and working environment. For example, invention patent CN202310283929.9 discloses a 360° righting device for fallen traffic cones. The device can adjust adaptively according to the falling direction of the traffic cone and then right the traffic cone, without relying on the forward or backward movement of the vehicle body to provide righting power. However, there are still the following shortcomings: (1) The position of the traffic cone needs to be accurately identified and determined, and then the retracting mechanism and the swing arm need to be moved to right the cone; (2) Even if the cone falls at a small angle, it is necessary to stop and right the cone, which seriously restricts the recovery efficiency; (3) After righting the traffic cone that falls in the direction of travel, it is still necessary to continue to increase the swing arm angle and rotate it to a non-blocking state before it can be recovered, which increases the operation time; (4) There is no method for identifying the fallen posture of the fallen traffic cone, nor how to right it according to the posture.

[0004] In addition, invention patent CN202410657491.0 discloses a method for intelligently identifying and locating traffic cones based on binocular vision, which is used to capture the situation of fallen traffic cones and identify their shape. However, there are still the following deficiencies: (1) It only identifies whether a traffic cone is in a fallen state, but does not perform targeted classification and identification of the state of the traffic cone; (2) It also does not provide a method for automatically righting and retracting fallen traffic cones.

[0005] In summary, there is an urgent need for a video-based automatic righting system for fallen traffic cones to solve the problems of accurate angle identification required for righting traffic cones, the problem of the mechanism interfering with the continued recovery of traffic cones after righting, and the problem of temporary parking required to recover traffic cones that have fallen in any direction. Summary of the Invention

[0006] The purpose of the present invention is to provide a video-based automatic righting system for fallen traffic cones, which solves the problem that the traffic cone righting requires precise angle identification and the problem that the mechanism interferes with the continued recovery of the traffic cone after righting.

[0007] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0008] A video-based automatic righting system for fallen traffic cones, including a recognition module, an information processing module, and a righting module;

[0009] The recognition module collects traffic cone videos through a monocular camera, uses the camera to obtain road images containing traffic cones in different states, defines the ROI range of the collected road images and performs perspective transformation for subsequent information processing;

[0010] The information processing module creates a dataset containing different traffic cone states and uses an improved YOLOv8 network to obtain the optimal pre-trained model. The pre-trained model detects and identifies the posture and angle of traffic cones, outputting the cone's posture and angle. It then determines whether to perform angle detection based on the cone's posture and angle. The cone's posture and angle are then used to classify the cone's state. The upright state is defined as the normal position of a cone, with the tip facing upward and the tail touching the ground. The different orientations of a cone in a collapsed state are defined as the front-righting state and the rear-righting state, respectively.

[0011] Improvements to the YOLOv8 network include: improving the backbone network by adding an improved Transformer structure at the end of the backbone network, retaining the global information of the feature map to a single axis through compression operations, and applying a self-attention mechanism to model long-range dependencies on each axis;

[0012] Improvements to the loss function: The algorithm involved includes the varifocal loss (VFL) function. By adjusting the weights of positive and negative samples, the model focuses more on difficult-to-classify samples during training, improving the detection accuracy of traffic cones in complex background environments.

[0013] The algorithms involved also include distribution focal loss (DFL) and ProbIoU, which are used to calculate the loss of the predicted bounding box and the true bounding box. DFL is used to regress the position of the predicted bounding box, and ProbIoU is used to calculate the degree of overlap between the predicted bounding box and the true bounding box, and jointly predict the position loss of the bounding box.

[0014] The straightening module includes a door-shaped bracket with two side guide arms and an upper crossbeam. The lower ends of the two side guide arms are each equipped with an arc-shaped guide plate, which is symmetrically installed. A first servo motor is installed in the center of the crossbeam to adjust the bracket deflection angle.

[0015] The guide arm is fixed in the middle with support arms extending forward and backward, with front and rear righting devices mounted on the ends of the support arms, respectively. Both the front and rear righting devices include a second servo motor and a shift lever. The shift lever is L-shaped, and the corresponding short arm end is coupled to the second servo motor through bevel gears to control the lowering or raising of the shift lever. The righting module can move up and down, with its maximum travel height slightly greater than the bottom length of the traffic cone, and its minimum travel height slightly above the ground.

[0016] According to the output information of the information processing module, the straightening module is controlled to rotate, straighten and straighten the traffic cone.

[0017] Furthermore, the recognition module collects the current road image in real time through the camera, performs grayscale, binarization and Gaussian blur processing on the image, and measures the distance S between the righting module and the edge of the traffic cone by drawing a histogram and using a sliding window algorithm;

[0018] Road acquisition is to use a camera to shoot road videos containing traffic cones in different states, save them to storage media via memory, and capture the videos frame by frame to obtain road images containing traffic cones under different road types, different environments, and different distance conditions.

[0019] A method for measuring the distance S between the righting module and the edge of the traffic cone by drawing a histogram and using a sliding window algorithm: based on the reference position of the traffic cone found by the histogram, a sliding window algorithm is used to identify the pixel where the traffic cone is located;

[0020] First, a sliding window is created. Starting from the bottom of the image, it slides upward within the image height range, searching for non-zero pixels within the window. The center position of the next window is relocated based on the pixels detected within the window. If the number of pixels detected in the current window exceeds a threshold, the center position of the next window is updated to the coordinates of the current window. This process is repeated until all windows of the image are processed, forming a curve showing the location of traffic cones.

[0021] According to the pixel coordinates of the traffic cones found, a quadratic polynomial is called to fit the traffic cones to obtain a quadratic polynomial curve model of the traffic cone edge. The offset between the pixel value of the traffic cone edge and the corresponding pixel value of the righting module is calculated and multiplied by the conversion coefficient to obtain the distance S between the righting module and the traffic cone edge.

[0022] Furthermore, the road image is demarcated into an ROI range, and the oblique-angle road image is transformed into a bird's-eye view to eliminate the perspective distortion in the original image; the formula is as follows:

[0023]

[0024] Where (x, y) is the point of the input road image, (x′, y′) is the point mapped to the output image after perspective transformation, w is the normalization factor of the homogeneous coordinates, and h is ij is the parameter to be found, and the solution is as follows:

[0025] Assume that the coordinates of the vertices of the ROI range border are (x R1 ,y R1 )、(x R2 ,y R2 )、(x R3 ,y R3 ) and (x R4 ,y R4 ), the output bird's-eye view border vertex coordinates are (x N1 ,y N1 )、(x N2 ,y N2 )、(x N3 ,y N3 ) and (x N4 ,y N4 ), the relationship between them is as follows:

[0026]

[0027] Substitute into h ij According to the posture and falling angle of traffic cones in the bird's-eye view, traffic cones are divided into three states: upright state, front-righting state and rear-righting state;

[0028] When the traffic cone is in a fallen state, use the smallest rectangular frame to select the top view of the traffic cone; based on the side view of the traffic cone in the upright state, set the coordinates of the upper left corner of the rectangle to (x1, y1) and the coordinates of the lower right corner to (x2, y2); use the coordinates of the center point of the rectangle to select the top view of the traffic cone; Establish a rectangular coordinate system for the origin; let the intersection of the tip of the traffic cone and the rectangular frame be (x3, y3). When y3≤0, the state is defined as the front righting state; when y3>0, the state is defined as the rear righting state.

[0029] Furthermore, for datasets of different traffic cone states, labelimg is used to annotate different traffic cone postures in the image, creating a dataset containing different traffic cone states and label files corresponding to the image information; the dataset is divided into training set, validation set, and test set; the training set images are passed into the improved YOLOv8 network in random order and trained through the processor; the network is evaluated using the test set, and the average accuracy mean of each test is recorded. When the training times reach the predetermined number, the training stops, and the network weight at the highest average accuracy mean is saved as the final weight.

[0030] Improved YOLOv8 network to obtain the best pre-training model method, assuming that the tensor input of the given image data is X∈R H×W×C , where R represents the batch size, H represents the image height, W represents the image width, and C represents the number of channels; Q, K, and V are obtained through three linear layers:

[0031]

[0032] Where W q , and is a learnable parameter;

[0033] In the W and H directions, horizontal compression and vertical compression are achieved by extracting the average values ​​of the Q, K, and V feature maps, respectively;

[0034]

[0035] Where, I W ∈R W×1 and I H ∈R H×1 are all-1 vectors of length W and H respectively;

[0036] Keep the global information on a single axis, then apply self-attention to model the long-term dependency of the corresponding axis separately, and finally add the two together through a broadcast operation:

[0037]

[0038] Finally, add a residual connection to Y and apply the weight matrix as The final output of 1×1Conv is:

[0039] Y out ∈R H×W×C =Conv 1×1 (Y+V)

[0040] Where,

[0041] Using an additional set of parameters and Applied to input X∈R H×W×C , and get the corresponding Q (e) , K (e) and V (e) :

[0042]

[0043] The three are spliced ​​in the channel dimension to model local spatial dependencies and enhance the perception of local details:

[0044]

[0045] Then use ReLU and 1×1Conv with BatchNorm to reduce the number of channels (2C qk +C v ) is mapped to C to generate detail enhancement features:

[0046] X qkv ∈R H×W×C =BatchNorm(Conv(ReLU(X qkv )))

[0047] Finally, the enhanced features are fused with the features given by the compression operation; by out The weights are generated by the Sigmoid function and then multiplied point by point with the enhanced features to obtain the fused output:

[0048] output∈R H×W×C =Sigmoid(Y out )⊙X qkv

[0049] The overall loss function can be expressed as:

[0050] Loss = λ cls L cls +λ dfl L dfl +λ box L box

[0051] Where L cls represents the classification loss, L dfl and L box Jointly represents the position loss of the predicted bounding box, λ cls ,λ dfl and λ box Represents the weights of each loss function, which are initially set to 0.5, 0.5, and 0.5 respectively;

[0052] VFL uses an adaptive weight adjustment method to adjust the weight of the sample according to the difficulty and weight distribution of the sample, which more effectively handles the problem of class imbalance; therefore, L cls It can be expressed as:

[0053]

[0054] Where p represents the predicted score, q represents the target score; α and γ are set to 1.25 and 2 by default;

[0055] DFL is used to regress and predict the position of the bounding box, and handles the regression problem of the bounding box through probability distribution. Therefore, L dfl It can be expressed as:

[0056] L dfl =DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 ))

[0057] Where y represents the true distance from the anchor point to an edge on the feature map size, y i is the largest integer less than y; y i+1 is the smallest integer greater than y, S i and S i+1 is the predicted probability of the corresponding position;

[0058] The ProbIoU calculation process is as follows:

[0059] First, the distances l, t, r, and b from the anchor point to the four edges are converted to x, y, w, and h, and then converted to Gaussian distribution according to the conversion equation. The mean of the obtained distribution is μ = (x, y) T ; The covariance matrix can be expressed as

[0060]

[0061] Where θ represents the rotation angle, W and H represent the width and height of the bounding box respectively; therefore, the mean and covariance matrices corresponding to the Gaussian distribution p of the true bounding box and the Gaussian distribution q of the predicted bounding box are expressed as:

[0062]

[0063] Bhattacharyya distance is used to calculate the difference between two Gaussian distributions and can be expressed as:

[0064]

[0065] The Bhattacharyya coefficient can be expressed as L box It can be expressed as:

[0066]

[0067] The final weights are loaded and the test set is evaluated to determine whether the network is overfitting. The training is completed when the test results are close to those in the test set.

[0068] Furthermore, the trained network model and final weights in the processor detect and identify the state of the traffic cone, and output the state of the traffic cone. The initial state of the righting module is at the minimum travel height, the initial state of the curved guide plate is parallel to the direction of vehicle movement, and the levers in the front and rear righting devices are both raised.

[0069] When the traffic cone is identified as being in an upright state, the vehicle continues to back up the distance S between the righting module and the edge of the traffic cone, so that the traffic cone passes through the righting module in an upright state, completing the traffic cone recovery operation;

[0070] When the traffic cone state is identified as the front righting state or the rear righting state, the improved YOLOv8 network is used to detect the deflection angle of the traffic cone and measure the deflection angle θ of the fallen traffic cone.

[0071] The deflection angle θ is defined by establishing a rectangular coordinate system with the upper left corner of the image as the origin; with the positive direction of the x-axis as 0°, the angle gradually increases as it rotates counterclockwise, and the angle value range is (-90°, 90°);

[0072] Use the smallest rectangle to select the traffic cone, select the endpoint A of the rectangle closest to the x-axis, and define its coordinates as (x A ,y A ), define the coordinates of the rectangle endpoint B, point C and point D in clockwise direction as (x B ,y B )、(x C ,y C ) and (x D ,y D );

[0073] Calculate the distance between point A and point B respectively and the distance between point A and point D When d1>d2, calculate the angle θ1 between the straight line where points A and B are located and the x-axis. The formula is as follows:

[0074]

[0075] At this time, the output angle θ=θ1;

[0076] When d1 < d2, calculate the angle θ2 between the straight line between point A and point D and the x-axis using the following formula:

[0077]

[0078] At this time, the output angle is θ=θ2.

[0079] When the deflection angle θ of the fallen traffic cone is small, the righting module maintains the minimum travel height, the vehicle continues to reverse, and the traffic cone is straightened; according to the state of the traffic cone, the operation process of the curved guide plate and the barrier is determined, and the vehicle continues to reverse to complete the traffic cone righting operation;

[0080] When the deflection angle θ of the fallen traffic cone is large, the righting module rises to its maximum travel height and determines the rotation direction and angle of the curved guide plate based on the detection result of the traffic cone's falling angle. The righting module descends to its minimum travel height, and the traffic cone enters the curved guide plate. Based on the state of the traffic cone and the deflection angle θ, the curved guide plate continues to rotate to complete the traffic cone straightening operation. The front / rear righting device controls the corresponding lever to be lowered, and the vehicle continues to reverse, completing the traffic cone righting operation.

[0081] After the righting is completed, the front / rear righting device controls the corresponding lever to lift up;

[0082] The distance S between the vehicle's reverse righting module and the edge of the traffic cone is set so that the righting module approaches the target traffic cone. The motor's rotation angle signal is output based on the state of the traffic cone and the angle of its fall, driving the servo motor to proceed to the next step.

[0083] When the deflection angle of the fallen traffic cone is |θ|≤θ′, where θ′ is 15°, the vehicle continues to reverse. The traffic cone with a smaller deflection angle is straightened under the guidance of the curved guide plate and the inertia of the vehicle. If the traffic cone is in the front righting state, the servo motor in the front righting device controls the lever to be lowered, and the vehicle continues to reverse, completing the traffic cone righting operation. If the traffic cone is in the rear righting state, the servo motor controls the bracket to drive the curved guide plate to rotate 180°, and the servo motor in the rear righting device controls the lever to be lowered. The vehicle continues to reverse, and the lever contacts the bottom of the traffic cone. Under the action of the vehicle's inertia, the traffic cone is straightened.

[0084] When the deflection angle of the fallen traffic cone |θ|>θ′, the righting module rises to the maximum travel height; if θ<0 at this time, the servo motor controls the bracket to drive the curved guide plate to rotate counterclockwise by 90°-|θ|; if θ>0 at this time, the servo motor controls the bracket to drive the curved guide plate to rotate clockwise by 90°-θ; the righting module moves down to the minimum travel height, allowing the traffic cone to completely enter the curved guide plate;

[0085] When the traffic cone is in the forward righting position,

[0086] If θ<0, the curved guide plate rotates 90°-|θ| clockwise to close the traffic cone;

[0087] If θ>0, the arc guide plate rotates counterclockwise 90°-θ to close the traffic cone;

[0088] The servo motor in the front straightening device controls the lever to be lowered, the vehicle continues to reverse, and the traffic cone is straightened;

[0089] When the traffic cone is in the rear-righting state,

[0090] If θ<0, the curved guide plate rotates 90°+|θ| counterclockwise to close the traffic cone;

[0091] If θ>0, the curved guide plate rotates 90°+θ clockwise to close the perpendicular traffic cone;

[0092] The servo motor in the rear straightening device controls the lever to be lowered, the vehicle continues to reverse, and the traffic cone is straightened;

[0093] After the righting is completed, the front / rear righting device controls the corresponding lever to lift up.

[0094] The present invention has the following beneficial effects: the present invention identifies the state and angle of traffic cones based on the righting method, integrates the rotation and righting function and the righting function into one, relaxes the requirements for the accuracy of traffic cone righting on the angle of fall through the righting mechanism design, solves the problem that the righting mechanism interferes with the continued recovery of traffic cones after the traffic cones are righted, and alleviates the problem that a traffic cone that has fallen in the righting state before recovery needs to be temporarily parked. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments.

[0096] Figure 1 : Flowchart of the traffic cone state recognition method of the present invention.

[0097] Figure 2 : Traffic cone state definition diagram of the present invention.

[0098] Figure 3 : This invention improves the YOLOv8 network structure.

[0099] Figure 4 : Traffic cone recognition angle definition diagram of the present invention.

[0100] Figure 5 : Schematic diagram of the righting module of the present invention.

[0101] Figure 6 : Schematic diagram of the front / rear righting device of the present invention.

[0102] In the accompanying drawings, the components represented by each reference numeral are listed as follows: bracket 2, arc-shaped guide rail plate 5, first servo motor 1, front straightening device 3, rear straightening device 4, second servo motor 3-3, gear lever 3-1, bevel gear 3-2. DETAILED DESCRIPTION

[0103] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0104] like Figure 1 Shown: A video-based automatic righting system for fallen traffic cones, including the following steps:

[0105] In step 1, use a camera to capture a video of a road containing traffic cones in different states and save it to a storage medium. Capture the video frame by frame to obtain images of the road containing traffic cones under different road types, different environments, and different distances. Then proceed to step 2.

[0106] Step 2: Define the ROI range of the collected road image and perform perspective transformation. Then go to step 3.

[0107] Sub-step 2.1: To eliminate the interference of environmental information such as the road and sky background and improve the real-time performance of the detection, the ROI range is defined for the collected road image.

[0108] In sub-step 2.2, the oblique road image is transformed into a bird's-eye view to eliminate the perspective distortion in the original image. The formula is as follows:

[0109]

[0110] Where (x, y) is the point of the input road image, (x′, y′) is the point mapped to the output image after perspective transformation, w is the normalization factor of the homogeneous coordinates, and h is ij is the parameter to be found, and the solution is as follows:

[0111] Assume that the coordinates of the vertices of the ROI range border are (x R1 ,y R1 )、(x R2 ,y R2 )、(x R3 ,y R3 ) and (x R4 ,y R4 ), the output bird's-eye view border vertex coordinates are (x N1 ,y N1 )、(x N2 ,y N2 )、(x N3 ,yN3 ) and (x N4 ,y N4 ), the relationship between them is as follows:

[0112]

[0113] Substitute into h ij .

[0114] Step 3: Based on the posture and falling angle of the traffic cones in the bird's-eye view, classify the traffic cones into three states: upright, front-righting, and rear-righting. Then go to step 4.

[0115] like Figure 2 As shown in sub-step 3.1, when the traffic cone is in a normal placement, the tip of the traffic cone is facing upward and the tail is grounded, and this state is defined as the upright state.

[0116] In sub-step 3.2, when the traffic cone is in the fallen state, use the smallest rectangle to select the top view of the traffic cone. Based on the side view of the traffic cone in the upright state, set the coordinates of the upper left corner of the rectangle to (x1, y1) and the coordinates of the lower right corner to (x2, y2). Establish a rectangular coordinate system for the origin. Let the intersection of the tip of the traffic cone and the rectangular frame be (x3, y3). When y3 ≤ 0, define this state as the front righting state; when y3 > 0, define this state as the rear righting state.

[0117] In step 4, create a dataset containing different traffic cone states and pre-train it using the improved YOLOv8 network to obtain the optimal model. Then proceed to step 5.

[0118] In sub-step 4.1, use labelimg to annotate the different postures of traffic cones in the image, creating a dataset containing different states of traffic cones and label files corresponding to the image information;

[0119] Sub-step 4.2, divide the dataset into training set, validation set and test set;

[0120] In substep 4.3, the training set images are randomly passed to the modified YOLOv8 network for training via the processor. The network is evaluated using the test set, and the mean average precision (APM) for each test is recorded. Training stops when the predetermined number of training runs is reached, and the network weights at the highest APM are saved as the final weights.

[0121] Furthermore, in step 4.3, the improvements to the YOLOv8 network include:

[0122] like Figure 3Figure 2 shows improvements to the backbone network. To more fully understand image content and improve the accuracy of traffic cone recognition, a Transformer structure is added to the end of the backbone network. Considering that this algorithm is deployed on mobile devices, the computational cost and memory consumption of the Transformer are very high, so the Transformer structure has been improved. A compression operation is used to retain the global information of the feature map to a single axis. The self-attention mechanism is then applied to each axis to model long-range dependencies, significantly reducing computational complexity. The specific steps are as follows:

[0123] Assume that the tensor input of the given image data is X∈R H×W×C , where R represents the batch size, H represents the image height, W represents the image width, and C represents the number of channels. Q, K, and V are obtained through three linear layers:

[0124]

[0125]

[0126] Where W q , and is a learnable parameter.

[0127] In the W and H directions, horizontal compression and vertical compression are achieved by extracting the average of the Q, K, and V feature maps, respectively.

[0128]

[0129] Where, I W ∈R W×1 and I H ∈R H×1 are all-1 vectors of length W and H respectively.

[0130] Keep the global information on a single axis, then apply self-attention to model the long-term dependency of the corresponding axis separately, and finally add the two together through a broadcast operation:

[0131]

[0132] Finally, add a residual connection to Y and apply the weight matrix as The final output of 1×1Conv is:

[0133] Y out ∈R H×W×C =Conv 1×1 (Y+V)

[0134] Where,

[0135] Since compression operation will lead to loss of detail information, convolution operation is used to enhance detail information. First, a set of additional parameters are used and Applied to input X∈R H×W×C , and get the corresponding Q (e) , K (e) and V (e) :

[0136]

[0137]

[0138] The three are spliced ​​in the channel dimension to model local spatial dependencies and enhance the perception of local details:

[0139]

[0140] Then use ReLU and 1×1Conv with BatchNorm to reduce the number of channels (2C qk +C v ) is mapped to C to generate detail enhancement features:

[0141] X qkv ∈R H×W×C =BatchNorm(Conv(ReLU(X qkv )))

[0142] Finally, the enhanced features are fused with the features given by the compression operation. out The weights are generated by the Sigmoid function and then multiplied point by point with the enhanced features to obtain the fused output:

[0143] output∈R H×W×C =Sigmoid(Y out )⊙X qkv

[0144] Improve the loss function. Considering that during the detection process, the background samples are far more than the traffic cone samples that need to be identified, the classification loss of this algorithm adopts the varifocal loss (VFL) function. This function adjusts the weights of positive and negative samples so that the model pays more attention to samples that are difficult to classify during training, thereby improving the detection accuracy of traffic cones in complex background environments. At the same time, the image of traffic cones is irregular in shape, and mutations, inconsistencies or instability may occur near certain specific boundaries or switching points, which in turn leads to increased angle sensitivity, high dependence on angle changes or inadaptability to angle changes, thereby affecting recognition accuracy. To this end, in order to better understand the feature information in the image, depict the offset of the predicted bounding box relative to the true bounding box, and avoid boundary discontinuity and angle sensitivity, this algorithm uses distribution focal loss (DFL) and ProbIoU to calculate the loss of the predicted bounding box and the true bounding box, where DFL is used to regress the position of the predicted bounding box and ProbIoU is used to calculate the degree of overlap between the predicted bounding box and the true bounding box. The overall loss function can be expressed as:

[0145] Loss = λ cls L cls +λ dfl L dfl +λ box L box

[0146] Where L cls represents the classification loss, L dfl and L box Jointly represents the position loss of the predicted bounding box, λ cls ,λ dfl and λ box represents the weights of each loss function, which are initially set to 0.5, 0.5, and 0.5 respectively.

[0147] VFL uses an adaptive weight adjustment method to adjust the weight of the sample according to the difficulty and weight distribution of the sample, which more effectively handles the problem of class imbalance. cls It can be expressed as:

[0148]

[0149] Where p represents the predicted score and q represents the target score. α and γ are set to 1.25 and 2 by default.

[0150] DFL is used to regress the predicted bounding box position. It handles the bounding box regression problem through probability distribution instead of directly regressing the bounding box coordinates. dfl It can be expressed as:

[0151] L dfl =DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 ))

[0152] Where y represents the true distance from the anchor point to an edge on the feature map size, y i Is the largest integer less than y. i+1 Is the smallest integer greater than y. i and S i+1 is the predicted probability of the corresponding position.

[0153] ProbIoU is used to calculate the degree of overlap between the predicted bounding box and the ground-truth bounding box and serves as a loss function to optimize bounding box regression. This function converts the predicted bounding box and the ground-truth bounding box into a two-dimensional Gaussian distribution and then uses the Bhattacharyya distance to calculate the difference between the two distributions. This method can avoid the shortcomings of traditional IoU (Intersection over Union) in terms of boundary discontinuity and angle sensitivity. The specific calculation process is as follows:

[0154] First, the distances l, t, r, and b from the anchor point to the four edges are converted to x, y, w, and h, and then converted to Gaussian distribution according to the conversion equation. The mean of the obtained distribution is μ = (x, y) T The covariance matrix can be expressed as

[0155]

[0156] Where θ represents the rotation angle, W and H represent the width and height of the bounding box respectively. Therefore, the mean and covariance matrices corresponding to the Gaussian distribution p of the true bounding box and the Gaussian distribution q of the predicted bounding box are expressed as:

[0157]

[0158] Bhattacharyya distance is used to calculate the difference between two Gaussian distributions and can be expressed as:

[0159]

[0160] The Bhattacharyya coefficient can be expressed as L box It can be expressed as:

[0161]

[0162] In sub-step 4.4, the final weights are loaded and the test set is evaluated to determine whether the network is overfitting. When the test results are close to those in the test set, the training is considered complete.

[0163] Step 5: During the vehicle reversing process, the camera collects the current road image in real time, processes the image, and measures the distance S between the righting module and the edge of the traffic cone.

[0164] In sub-step 5.1, to eliminate redundant color information in the image and retain only the brightness difference information between the traffic cone image and the surrounding road area for subsequent rapid detection tasks, the bird's-eye view image is converted from the BGR color space to the HLS color space and grayscaled.

[0165] In sub-step 5.2, to highlight the edges of traffic cones in the image, the grayscale image is binarized and blurred using Gaussian blur. The noise and details in the image are reduced by taking a weighted average of the surrounding pixels of each pixel.

[0166] In substep 5.3, to determine the location of the pixels where the traffic cones are located, a histogram is plotted on the preprocessed image. The histogram shows the distribution of the accumulated values ​​of white pixels at different vertical positions in the image. The peak of the histogram closest to the bottom of the image is the reference position of the traffic cone edge.

[0167] In substep 5.4, based on the cone's baseline position found in the histogram, a sliding window algorithm is used to identify the pixels where the cone is located. First, create a sliding window and, starting at the bottom of the image, slide it upward within the image height, searching for nonzero pixels within the window. Based on the pixels detected within the window, the center position of the next window is relocated. If the number of pixels detected within the current window exceeds a threshold, the center position of the next window is updated to the coordinates of the current window. Repeat this process until all windows of the image have been processed, forming a curve that represents the locations of the cones.

[0168] In substep 5.5, based on the found pixel coordinates of the traffic cone, a quadratic polynomial is applied to the cone to obtain a quadratic polynomial curve model of the cone edge. The offset between the pixel values ​​at the cone edge and the corresponding pixel values ​​of the righting module is calculated and multiplied by the conversion factor to obtain the distance S between the righting module and the cone edge.

[0169] In step 6, the processor uses the trained network model and final weights to detect and identify the traffic cone state, outputting the cone's state. The righting module is initially at its minimum travel height, the curved guide plate is initially parallel to the vehicle's direction of motion, and the front and rear righting levers are both raised.

[0170] When the traffic cone is identified as being in an upright state, the vehicle continues to reverse the distance S between the righting module and the edge of the traffic cone, so that the traffic cone passes through the righting module in an upright state, completing the traffic cone recovery operation and proceeding to step 5;

[0171] like Figure 4 As shown: When the traffic cone state is identified as the front righting state or the rear righting state, the improved YOLOv8 network is used to detect the deflection angle of the traffic cone, the deflection angle θ of the fallen traffic cone is measured, and step 7 is executed.

[0172] Sub-step 6.1, define the rotation angle. Establish a rectangular coordinate system with the upper left corner of the image as the origin. With the positive x-axis direction at 0°, rotate counterclockwise, gradually increasing the angle within the range of (-90°, 90°).

[0173] Sub-step 6.2, select the traffic cone with the smallest rectangle, select the endpoint A of the rectangle closest to the x-axis, and define its coordinates as (x A ,y A ), define the coordinates of the rectangle endpoint B, point C and point D in clockwise direction as (x B ,y B )、(x C ,y C ) and (x D ,y D ).

[0174] Sub-step 6.2, calculate the distance between point A and point B respectively and the distance between point A and point D When d1>d2, calculate the angle θ1 between the straight line where points A and B are located and the x-axis. The formula is as follows:

[0175]

[0176] At this time, the output angle θ=θ1;

[0177] When d1 < d2, calculate the angle θ2 between the straight line between point A and point D and the x-axis using the following formula:

[0178]

[0179] At this time, the output angle is θ=θ2.

[0180] Step 7: When the deflection angle θ of the fallen traffic cone is small, the righting module maintains its minimum travel height, and the vehicle continues to reverse, straightening the cone. The curved guide plate and barrier lever operation process is determined based on the cone's current state. The vehicle continues to reverse, completing the cone righting operation.

[0181] When the cone reaches a significant deflection angle θ, the righting module rises to its maximum travel height. Based on the cone's angle of fall, the curved guide plate's rotation direction and angle are determined. The righting module then descends to its minimum travel height, allowing the cone to enter the curved guide plate. Based on the cone's current state and deflection angle θ, the curved guide plate continues to rotate, completing the cone's straightening. The front / rear righting device lowers the corresponding lever, allowing the vehicle to continue reversing, completing the cone righting operation.

[0182] After the righting is completed, the front / rear righting device controls the corresponding lever to lift up, and then go to step 5.

[0183] In sub-step 7.1, the vehicle reverses the distance S between the righting module and the edge of the traffic cone, bringing the righting module close to the target traffic cone. Based on the state of the traffic cone and the angle of its fall, a motor rotation signal is output to drive the servo motor to proceed to the next step.

[0184] Sub-step 7.2: When the deflection angle of the fallen traffic cone is |θ| ≤ θ′ (θ′ is generally 15°), the vehicle continues to reverse, and the traffic cone with a smaller deflection angle is straightened under the guidance of the curved guide plate and the inertia of the vehicle. If the traffic cone is in the front straightening state, the servo motor in the front straightening device controls the lever to be lowered, and the vehicle continues to reverse, completing the traffic cone straightening operation. If the traffic cone is in the rear straightening state, the servo motor controls the bracket to drive the curved guide plate to rotate 180°, and the servo motor in the rear straightening device controls the lever to be lowered. The vehicle continues to reverse, and the lever contacts the bottom of the traffic cone. Under the action of vehicle inertia, the traffic cone is straightened.

[0185] When the deflection angle of the fallen traffic cone exceeds θ′, the straightening module rises to its maximum travel height. If θ is less than 0, the servo motor controls the bracket to rotate the curved guide plate counterclockwise by 90° - |θ|. If θ is greater than 0, the servo motor controls the bracket to rotate the curved guide plate clockwise by 90° - θ. The straightening module then descends to its minimum travel height, allowing the cone to fully enter the curved guide plate.

[0186] When the traffic cone is in the forward righting position,

[0187] If θ<0, the curved guide plate rotates 90°-|θ| clockwise to close the traffic cone;

[0188] If θ>0, the arc guide plate rotates counterclockwise 90°-θ to close the traffic cone.

[0189] The servo motor in the front righting device controls the lever to be lowered. The vehicle continues to reverse and the traffic cone is straightened.

[0190] When the traffic cone is in the rear-righting state,

[0191] If θ<0, the curved guide plate rotates 90°+|θ| counterclockwise to close the traffic cone;

[0192] If θ>0, the arc guide plate rotates 90°+θ in clockwise direction to close the traffic cone.

[0193] The servo motor in the rear righting device controls the lever to be lowered. The vehicle continues to reverse and the traffic cone is straightened.

[0194] Sub-step 7.3: After the righting is completed, the front / rear righting device controls the corresponding lever to lift up.

[0195] like Figure 5-Figure 6 As shown: the straightening module includes a door-shaped bracket 2, with two side guide arms and an upper side crossbeam, the lower ends of the two side guide arms are each installed with an arc-shaped guide plate 5, the arc-shaped guide plate 5 is symmetrically installed, and a first servo motor 1 is installed in the center of the crossbeam for adjusting the deflection angle of the bracket 2;

[0196] The middle part of the guide arm is fixed with support arms extending forward and backward, and the ends of the support arms are respectively installed with a front straightening device 3 and a rear straightening device 4; the front straightening device 3 and the rear straightening device 4 both include a second servo motor 3-3 and a gear lever 3-1, and the gear lever 3-1 is L-shaped, and the corresponding short arm end is transmitted with the second servo motor 3-3 through a bevel gear 3-2, which is used to control the gear lever 3-1 to be lowered or raised.

[0197] This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention.

Claims

1. Video-based automatic righting system for fallen traffic cones, characterized by: It includes an identification module, an information processing module and a righting module; The recognition module uses a monocular camera to capture traffic cone videos, obtain road images containing traffic cones in different states, define the ROI range of the collected road images, and perform perspective transformation for subsequent information processing; The information processing module creates a dataset containing different traffic cone states and uses an improved YOLOv8 network to obtain the optimal pre-trained model. The pre-trained model detects and identifies the posture and angle of traffic cones, and outputs the cone's state and angle. Improvements to the YOLOv8 network include: improving the backbone network by adding an improved Transformer structure at the end of the backbone network, retaining the global information of the feature map to a single axis through compression operations, and applying a self-attention mechanism to model long-range dependencies on each axis; Improvements to the loss function: The algorithms involved include the varifocal loss (VFL) function, which improves the detection accuracy of traffic cones in complex background environments by adjusting the weights of positive and negative samples. The algorithms involved also include the distribution focal loss (DFL) and ProbIoU, which are used to calculate the loss between the predicted bounding box and the true bounding box. DFL is used to regress the position of the predicted bounding box, and ProbIoU is used to calculate the degree of overlap between the predicted bounding box and the true bounding box, jointly predicting the position loss of the bounding box. According to the output information of the information processing module, the straightening module is controlled to rotate, straighten and straighten the traffic cone.

2. The video-based automatic righting system for fallen traffic cones according to claim 1, characterized in that: The straightening module comprises a door-shaped bracket (2) having two side guide arms and an upper side crossbeam, wherein the lower ends of the two side guide arms are both provided with arc-shaped guide rail plates (5), the arc-shaped guide rail plates (5) are symmetrically provided, and a first servo motor (1) is provided in the center of the crossbeam for adjusting the deflection angle of the bracket (2); The middle part of the guide arm is fixed with support arms extending forward and backward, and the ends of the support arms are respectively installed with a front righting device (3) and a rear righting device (4); the front righting device (3) and the rear righting device (4) both include a second servo motor (3-3) and a gear lever (3-1); the gear lever (3-1) is L-shaped, and the corresponding short arm end is driven by the second servo motor (3-3) through a bevel gear (3-2) to control the gear lever (3-1) to be lowered or raised; the righting module can move up and down, and its maximum travel height from the ground is slightly greater than the bottom length of the traffic cone, and the minimum travel height is slightly above the ground.

3. The video-based automatic righting system for fallen traffic cones according to claim 1, characterized in that: The recognition module collects the current road image in real time through the camera, performs grayscale, binarization and Gaussian blur processing on the image, and measures the distance S between the righting module and the edge of the traffic cone by drawing a histogram and using a sliding window algorithm; Road acquisition is to use a camera to shoot road videos containing traffic cones in different states, save them to storage media via memory, and capture the videos frame by frame to obtain road images containing traffic cones under different road types, different environments, and different distance conditions.

4. The video-based automatic righting system for fallen traffic cones according to claim 3, characterized in that: A method for measuring the distance S between the righting module and the edge of the traffic cone by drawing a histogram and using a sliding window algorithm: based on the reference position of the traffic cone found by the histogram, a sliding window algorithm is used to identify the pixel where the traffic cone is located; First, create a sliding window, starting from the bottom of the image, and slide the window upward within the image height range. Search for non-zero pixels in the window, and reposition the center position of the next window based on the pixels detected in the window. If the number of pixels detected in the current window exceeds the threshold, the center position of the next window is updated to the coordinates of the current window. The above process is repeated until all windows of the image are processed, forming a curve where the traffic cones are located. According to the pixel coordinates of the traffic cones found, a quadratic polynomial is called to fit the traffic cones to obtain a quadratic polynomial curve model of the traffic cone edge. The offset between the pixel value of the traffic cone edge and the corresponding pixel value of the righting module is calculated and multiplied by the conversion coefficient to obtain the distance S between the righting module and the traffic cone edge.

5. The video-based automatic righting system for fallen traffic cones according to claim 1, characterized in that: The ROI range is defined on the road image, and the oblique-angle road image is transformed into a bird's-eye view to eliminate the perspective distortion in the original image. The formula is as follows: Where (x, y) is the point of the input road image, (x′, y′) is the point mapped to the output image after perspective transformation, w is the normalization factor of the homogeneous coordinates, and h is ij is the parameter to be found, and the solution is as follows: Assume that the coordinates of the vertices of the ROI range border are (x R1 ,y R1 )、(x R2 ,y R2 )、(x R3 ,y R3 ) and (x R4 ,y R4 ), the output bird's-eye view border vertex coordinates are (x N1 ,y N1 )、(x N2 ,y N2 )、(x N3 ,y N3 ) and (x N4 ,y N4 ), the relationship between them is as follows: Substitute into h ij According to the posture and falling angle of traffic cones in the bird's-eye view, traffic cones are divided into three states: upright state, front-righting state and rear-righting state; When the traffic cone is in a fallen state, use the smallest rectangular frame to select the top view of the traffic cone; based on the side view of the traffic cone in the upright state, set the coordinates of the upper left corner of the rectangle to (x1, y1) and the coordinates of the lower right corner to (x2, y2); use the coordinates of the center point of the rectangle to select the top view of the traffic cone; Establish a rectangular coordinate system for the origin; let the intersection of the tip of the traffic cone and the rectangular frame be (x3, y3). When y3≤0, the state is defined as the front righting state; when y3>0, the state is defined as the rear righting state.

6. The video-based automatic righting system for fallen traffic cones according to claim 5, characterized in that: A dataset of different traffic cone states is created. Labelimg is used to annotate the different traffic cone postures in the image, creating a dataset containing different traffic cone states and label files corresponding to the image information. The dataset is divided into training, validation, and test sets. The training set images are passed into the improved YOLOv8 network in random order and trained by the processor. The network is evaluated using the test set, and the average precision of each test is recorded. When the training reaches the predetermined number of times, the training stops, and the network weight at the highest average precision is saved as the final weight.

7. The video-based automatic righting system for fallen traffic cones according to claim 6, characterized in that: Improved YOLOv8 network to obtain the best pre-training model method, assuming that the tensor input of the given image data is X∈R H×W×C , where R represents the batch size, H represents the image height, W represents the image width, and C represents the number of channels; Q, K, and V are obtained through three linear layers: Where, and is a learnable parameter; In the W and H directions, horizontal compression and vertical compression are achieved by extracting the average values ​​of the Q, K, and V feature maps, respectively; Where, I W ∈R W×1 and I H ∈R H×1 are all-1 vectors of length W and H respectively; Keep the global information on a single axis, then apply self-attention to model the long-term dependency of the corresponding axis separately, and finally add the two together through a broadcast operation: Finally, add a residual connection to Y and apply the weight matrix as The final output of 1×1Conv is: AND out ∈R H×W×C =Conv 1×1 (Y+V) Where, Using an additional set of parameters and Applied to input X∈R H×W×C , and get the corresponding Q (e) , K (e) and V (e) : The three are spliced ​​in the channel dimension to model local spatial dependencies and enhance the perception of local details: Then use ReLU and 1×1Conv with BatchNorm to reduce the number of channels (2C qk +C v ) is mapped to C to generate detail enhancement features: X qkv ∈R H×W×C =BatchNorm(Conv(ReLU(X qkv ))) Finally, the enhanced features are fused with the features given by the compression operation; by out The weights are generated by the Sigmoid function and then multiplied point by point with the enhanced features to obtain the fused output: output∈R H×W×C =Sigmoid(Y out )⊙X qkv The overall loss function can be expressed as: Loss=λ cls L cls +λ dfl L dfl +λ box L box Where L cls represents the classification loss, L dfl and L box Jointly represents the position loss of the predicted bounding box, λ cls ,λ dfl and λ box Represents the weights of each loss function, which are initially set to 0.5, 0.5, and 0.5 respectively; VFL uses an adaptive weight adjustment method to adjust the weight of the sample according to the difficulty and weight distribution of the sample, which more effectively handles the problem of class imbalance; therefore, L cls It can be expressed as: Where p represents the predicted score, q represents the target score; α and γ are set to 1.25 and 2 by default; DFL is used to regress and predict the position of the bounding box, and handles the regression problem of the bounding box through probability distribution. Therefore, L dfl It can be expressed as: L dfl =DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(y-y i )log(S i+1 )) Where y represents the true distance from the anchor point to an edge on the feature map size, y i is the largest integer less than y; y i+1 is the smallest integer greater than y, S i and S i+1 is the predicted probability of the corresponding position; The ProbIoU calculation process is as follows: First, the distances l, t, r, and b from the anchor point to the four edges are converted to x, y, w, and h, and then converted to Gaussian distribution according to the conversion equation. The mean of the obtained distribution is μ = (x, y) T ; The covariance matrix can be expressed as Where θ represents the rotation angle, W and H represent the width and height of the bounding box respectively; therefore, the mean and covariance matrices corresponding to the Gaussian distribution p of the true bounding box and the Gaussian distribution q of the predicted bounding box are expressed as: m p =(x p ,y p ), m q =(x q ,y q ), Bhattacharyya distance is used to calculate the difference between two Gaussian distributions and can be expressed as: The Bhattacharyya coefficient can be expressed as L box It can be expressed as: The final weights are loaded and the test set is evaluated to determine whether the network is overfitting. The training is completed when the test results are close to those in the test set.

8. The video-based automatic righting system for fallen traffic cones according to claim 2, characterized in that: The processor uses the trained network model and final weights to detect and identify the state of traffic cones and output the cone's state. The initial state of the righting module is at the minimum travel height, the initial state of the curved guide plate is parallel to the direction of vehicle movement, and the levers in the front and rear righting devices are both raised. When the traffic cone is identified as being in an upright state, the vehicle continues to back up the distance S between the righting module and the edge of the traffic cone, so that the traffic cone passes through the righting module in an upright state, completing the traffic cone recovery operation; When the traffic cone state is identified as the front righting state or the rear righting state, the improved YOLOv8 network is used to detect the deflection angle of the traffic cone and measure the deflection angle θ of the fallen traffic cone.

9. The video-based automatic righting system for fallen traffic cones according to claim 8, characterized in that: The deflection angle θ is defined by establishing a rectangular coordinate system with the upper left corner of the image as the origin; with the positive direction of the x-axis as 0°, the angle gradually increases as it rotates counterclockwise, and the angle value range is (-90°, 90°); Use the smallest rectangle to select the traffic cone, select the endpoint A of the rectangle closest to the x-axis, and define its coordinates as (x A ,y A ), define the coordinates of the rectangle endpoint B, point C and point D in clockwise direction as (x B ,y B )、(x C ,y C ) and (x D ,y D ); Calculate the distance between point A and point B respectively and the distance between point A and point D When d1>d2, calculate the angle θ1 between the straight line where points A and B are located and the x-axis. The formula is as follows: At this time, the output angle θ=θ1; When d1 < d2, calculate the angle θ2 between the straight line between point A and point D and the x-axis using the following formula: At this time, the output angle is θ=θ2.

10. The video-based automatic righting system for fallen traffic cones according to claim 9, characterized in that: When the deflection angle θ of the fallen traffic cone is small, the righting module maintains the minimum travel height, the vehicle continues to reverse, and the traffic cone is straightened; according to the state of the traffic cone, the operation process of the curved guide plate and the barrier is determined, and the vehicle continues to reverse to complete the traffic cone righting operation; When the deflection angle θ of the fallen traffic cone is large, the righting module rises to its maximum travel height and determines the rotation direction and angle of the curved guide plate based on the detection result of the traffic cone's falling angle. The righting module descends to its minimum travel height, and the traffic cone enters the curved guide plate. Based on the state of the traffic cone and the deflection angle θ, the curved guide plate continues to rotate to complete the traffic cone straightening operation. The front / rear righting device controls the corresponding lever to be lowered, and the vehicle continues to reverse, completing the traffic cone righting operation. After the righting is completed, the front / rear righting device controls the corresponding lever to lift up; The distance S between the vehicle's reverse righting module and the edge of the traffic cone is set so that the righting module approaches the target traffic cone. The motor's rotation angle signal is output based on the state of the traffic cone and the angle of its fall, driving the servo motor to proceed to the next step. When the deflection angle of the fallen traffic cone is |θ|≤θ′, where θ′ is 15°, the vehicle continues to reverse. The traffic cone with a smaller deflection angle is straightened under the guidance of the curved guide plate and the inertia of the vehicle. If the traffic cone is in the front righting state, the servo motor in the front righting device controls the lever to be lowered, and the vehicle continues to reverse, completing the traffic cone righting operation. If the traffic cone is in the rear righting state, the servo motor controls the bracket to drive the curved guide plate to rotate 180°, and the servo motor in the rear righting device controls the lever to be lowered. The vehicle continues to reverse, and the lever contacts the bottom of the traffic cone. Under the action of the vehicle's inertia, the traffic cone is straightened. When the deflection angle of the fallen traffic cone |θ|>θ′, the righting module rises to the maximum stroke height; if θ<0 at this time, the first servo motor controls the bracket to drive the curved guide plate to rotate counterclockwise by 90°-|θ|; if θ>0 at this time, the first servo motor controls the bracket to drive the curved guide plate to rotate clockwise by 90°-θ; the righting module moves down to the minimum stroke height, allowing the traffic cone to completely enter the curved guide plate; When the traffic cone is in the forward righting position, If θ<0, the curved guide plate rotates 90°-|θ| clockwise to close the traffic cone; If θ>0, the arc guide plate rotates counterclockwise 90°-θ to close the traffic cone; The servo motor in the front straightening device controls the lever to be lowered, the vehicle continues to reverse, and the traffic cone is straightened; When the traffic cone is in the rear-righting state, If θ<0, the curved guide plate rotates 90°+|θ| counterclockwise to close the traffic cone; If θ>0, the curved guide plate rotates 90°+θ clockwise to close the perpendicular traffic cone; The servo motor in the rear straightening device controls the lever to be lowered, the vehicle continues to reverse, and the traffic cone is straightened; After the righting is completed, the front / rear righting device controls the corresponding lever to lift up.

Images