Perception Data Optimization Method Based on Autonomous UAV Applications

Abstract

The present invention discloses a method for optimizing perception data based on the application of autonomous drones, belonging to the technical field of artificial intelligence. It includes step S100 of traversing data according to the structure of a dynamic array; step S200 of setting two arrays to track data, where one array is an actual value array for recording input data, and the other array is a modified value array for recording data after correcting newly incoming data; step S300 of comparing the data of the two arrays, and selecting the data corresponding to the frame numbers of different arrays for output according to the offset situation and confidence level of the data corresponding to the frame numbers in the two arrays. By comparing and tracking data between the two arrays and outputting optimized data, the present invention can solve the downstream response problem caused by the jump or interruption of the original data collected when the drone body is flying at high speed.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and more specifically, relates to a method for optimizing perception data based on the application of autonomous unmanned aerial vehicles (UAVs). Background Art

[0002] When UAVs perform tasks, they usually encounter relatively complex perception problems: for example, the inherent error distribution of on-board sensors changes in different scenarios; the data collected during the high-speed flight of the airframe undergoes lag drift; target data is missed or misdetected, resulting in jumps, etc. These problems will cause a large error between the calculation result of the collected data and the expected processing result. Since perception data is strong-demand information when UAVs perform autonomous tasks, it cannot drift or be discarded for a long time, otherwise the UAV task will fail and manual intervention is required. Therefore, it is very important to optimize the data processing.

[0003] Current algorithms for data optimization processing include Kalman filter series, motion compensation algorithms, Gauss-Newton series nonlinear optimization algorithms, etc. Taking the Chinese invention patent with the application number CN201910708585.5, the application date of August 1, 2019, and the name of a GPS / INS integrated navigation and positioning method based on minimum upper bound filtering as an example for the Kalman filter series data optimization processing algorithm, the application includes the following steps: Step 1, navigation information acquisition stage; Step 2, navigation model establishment stage; Step 3, navigation error judgment stage; Step 4, navigation information correction; Step 5, optimal state estimation information output. The inventive method first models the malfunction deviation and interference parts in the GPS measurement signal as unknown input items in the sensor measurement, and on this basis, designs a GPS / INS integrated navigation and positioning method based on minimum upper bound filtering. This method can effectively suppress the positioning error and achieve accurate self-positioning and navigation when the GPS signal is disturbed or fails for a short time. Specifically, free parameters are introduced into the position estimation variance in the integrated navigation filtering algorithm, and an upper bound expression of the position estimation variance is constructed. The optimized value of its estimation variance is obtained through real-time convex optimization dynamic optimization, and accurate navigation information is output, thereby ensuring the stable operation of the integrated navigation system. However, it also has the following deficiencies: As a relatively common fusion algorithm, the Kalman series filter has good convergence in the backend optimization when the data is valid, but when dealing with relatively complex or input singularities, the response will show sudden errors, which is not safe enough for UAV applications.

[0004] For the motion compensation algorithm, taking the Chinese invention patent with the application number CN201410833820.9, the application date of December 26, 2014, and the title of an automatic reconnaissance system for an unmanned aerial vehicle (UAV)-borne optoelectronic stabilization platform based on image matching as an example, in this application, the path planning and automatic scanning control of automatic scanning are carried out through the image registration relationship of the optoelectronic payload under large and small fields of view, and the automatic traversal scanning of the preset observation area image is completed; during automatic scanning, image matching is performed on adjacent image sequences, motion compensation is performed through the matching relationship, and inter-frame difference operation is performed to detect moving targets, realizing the real-time automatic recognition of moving targets in the image and improving the reconnaissance accuracy and efficiency. However, it also has the following deficiencies: The motion compensation algorithm has a good compensation effect on data during uniform motion. However, there are speed jump points and many cases of rapid acceleration and deceleration during the flight of the UAV. At this time, if the motion compensation algorithm is used, the result will have a large deviation and does not meet the requirements.

[0005] For the Gauss-Newton series of nonlinear optimization algorithms, taking the Chinese invention patent with the application number CN201710332021.7, the application date of May 12, 2017, and the title of a data processing method for remote scanning based on a UAV platform as an example, this invention uses a UAV as a measurement platform, and there are significant differences in the real-time correction and compensation of data during the measurement process compared with the prior art. The data fusion method proposed in this invention not only solves the problem of the movement of the measurement platform, but also compensates for the platform rotation and vibration during the measurement process, and solves the spatial scanning during the remote movement process. The rich scanning results have important practical significance for application fields such as cave research, tunnel rescue, and the reconstruction of disaster sites such as earthquakes / fires. The complete, accurate, and rich three-dimensional results have important application significance. However, it also has the following deficiencies: Although the optimal likelihood estimation is performed on the system, its operation efficiency and solution efficiency are relatively low. Summary of the Invention

[0006] 1. Problems to be Solved

[0007] In view of the problems existing in the prior art, the present invention provides a method and system for optimizing perception data based on autonomous UAV applications, which optimizes the backend processing results of the data collected by the onboard sensors to solve the downstream response problems caused by the jump or interruption of the original data.

[0008] 2. Technical Solutions

[0009] To solve the above problems, the present invention adopts the following technical solutions.

[0010] The method for optimizing perception data based on autonomous UAV applications includes the following steps:

[0011] Step S100: Traverse the data according to the structure of the dynamic array;

[0012] Step S200: Set two arrays to track the data. One array is the actual value array for recording the input data, and the other array is the modified value array for recording the data after correcting the newly incoming data;

[0013] Step S300: Compare the data in the two arrays, and select the data corresponding to the frame numbers in different arrays for output according to the offset situation and confidence level of the data corresponding to the frame numbers in the two arrays.

[0014] Its preferred technical solution is:

[0015] In the method for optimizing perception data based on autonomous UAV applications as described above, in step S200, the data recorded in the modified value array after correcting the newly incoming data specifically includes:

[0016] Step S210: Determine whether the degree of data offset of the newly incoming data is within the preset data offset range. If not, enter step S220; if so, enter step S230;

[0017] Step S220: Optimize the data with data offset, correct the data according to the degree of data offset, and record the corrected data in the modified value array;

[0018] Step S230: Record the newly incoming data in the modified value array.

[0019] In the method for optimizing perception data based on autonomous UAV applications as described above, in step S220, optimizing the data with data offset, correcting the data according to the degree of data offset, and recording the corrected data in the modified value array specifically includes:

[0020] If the newly incoming data has a large offset, compensate the data and add it to the modified value array;

[0021] If the newly incoming data has a small offset, smooth the data and add it to the modified value array.

[0022] In the method for optimizing perception data based on autonomous UAV applications as described above, if the newly incoming data has a large offset, directly compensating the result and adding it to the modified value array specifically includes:

[0023] Step S211: Take the minimum second moment of the previous several frames of data, use this data to solve and output the current frame data, and add the output result to the modified value array.

[0024] In the method for optimizing perception data based on autonomous UAV applications as described above, if the newly incoming data has a small offset, smooth the data and add it to the modified value array specifically includes:

[0025] Step S212: Estimate the data using a state estimation method based on the filtering idea, replace the input value with the estimated value and output it, and add the output result to the modified value group;

[0026] Step S213: Add the current gain result to the buffer, update it if the results are continuous, and clear the result after a certain number of frames if they are not continuous.

[0027] In the perception data optimization method based on autonomous UAV applications as described above, in step S300, when comparing the data of two arrays, according to the offset situation of the data corresponding to the frame numbers in the two arrays and the confidence levels, selecting the data corresponding to the frame numbers in different arrays for output specifically includes:

[0028] Step S310: Compare the data in the two arrays frame by frame to determine whether the values of the data corresponding to the frame numbers are the same;

[0029] Step S320: Determine the confidence evaluation level of the data of the current frame number in the actual value group;

[0030] Step S330: Select the data corresponding to the frame numbers in different arrays as the output data according to whether the values corresponding to the frame numbers are the same and the output confidence levels of different arrays.

[0031] In the perception data optimization method based on autonomous UAV applications as described above, in step S320, determining the confidence evaluation level of the data of the current frame number in the actual value group specifically includes:

[0032] Step S321: Establish the confidence evaluation level of the newly entered data, and the confidence evaluation level of the data of the current frame number in the actual value group includes at least low, relatively low, medium, relatively high, and high;

[0033] Step S322: Determine the magnitude of the confidence of the newly entered data in the actual value group;

[0034] Step S323: Determine the confidence evaluation level of the newly entered data according to the magnitude of the confidence of the newly entered data in the actual value group.

[0035] In the perception data optimization method based on autonomous UAV applications as described above, in step S330, selecting the data corresponding to the frame numbers in different arrays as the output data according to whether the values corresponding to the frame numbers are the same and the output confidence levels of different arrays specifically includes:

[0036] If the values of the data of the current frame number in the actual value group and the modified value group are the same, and the output confidence level of the data of the current frame number in the actual value group is high or relatively high, select the data of the current frame number in the actual value group as the output data;

[0037] If the values of the current frame number data in the actual value group are the same as those in the modified value group, and the output confidence level of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the actual value group as the output data.

[0038] In the perception data optimization method based on autonomous UAV applications as described above, in step S330, selecting the data of the corresponding frame number in different arrays as the output data according to whether the values of the corresponding frame numbers are the same and the output confidence levels of different arrays further includes:

[0039] If the values of the current frame number data in the actual value group are different from those in the modified value group, determine the magnitude of the difference between the values of the current frame number data in the actual value group and the modified value group, where:

[0040] If the values of the current frame number data in the actual value group are different from those in the modified value group and the difference is small, and at the same time the output confidence level of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group as the output data;

[0041] If the values of the current frame number data in the actual value group are different from those in the modified value group and the difference is large, and at the same time the output confidence level of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the modified value group as the output data;

[0042] If the values of the current frame number data in the actual value group are different from those in the modified value group and the difference is large, and at the same time the output confidence level of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group or the modified value group as the output data according to the situation;

[0043] If the values of the current frame number data in the actual value group are different from those in the modified value group and the difference is large, and at the same time the output confidence level of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the modified value group as the output data.

[0044] In the perception data optimization method based on autonomous UAV applications as described above, if the values of the current frame number data in the actual value group are different from those in the modified value group and the difference is large, and at the same time the output confidence level of the current frame number data in the actual value group is high or at a relatively high level, selecting the data of the current frame number in the actual value group or the modified value group as the output data according to the situation specifically includes:

[0045] Calculate the root mean square error RMSE for the data of the previous several frames;

[0046] If the root mean square error RMSE of the data of the previous several frames is less than or equal to the preset root mean square error, select the data of the current frame number in the actual value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to a relatively high level, and delete the data of the current frame number in the modified value group;

[0047] If the root mean square error (RMSE) of the previous several frames of data is greater than the preset root mean square error, determine the confidence level of the previous several frames of data; where:

[0048] If the confidence levels of the previous several frames of data are all at a high or relatively high level, select the data of the current frame number in the actual value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to medium; otherwise:

[0049] Select the data of the current frame number in the modified value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to a relatively low level.

[0050] 3. Beneficial Effects

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] (1) The present invention tracks data by setting two dynamic arrays. The two dynamic arrays are divided into an actual value group and a modified value group. Among them, the actual value group only records the input data. When new data comes in, the previous several frames of data will be estimated to correct the new data; if the new data has a large deviation, the result will be directly compensated and added to the modified value group. If the new data has a small deviation, the data will be smoothed and added to the modified value group; if the data is not within the set optimization range, no change will be made. By comparing and tracking data between the two arrays and outputting optimized data, the present invention can solve the downstream response problem caused by the jump or interruption of the original data collected during the high-speed flight of the UAV body;

[0053] (2) The present invention selects the data of the corresponding frame number in different arrays as the output data according to whether the values corresponding to the frame numbers are the same and the different output confidence levels of different arrays. At the same time, with the help of the principle and idea of fuzzy mathematics, the problem of whether it is reasonable to use the corrected data as the subsequent frame optimization step after focusing on correcting the new data is considered, so that the data result output by the perception data optimization method based on the autonomous UAV application in this embodiment is more reliable, and the spatio-temporal efficiency and accuracy are higher. Description of the Drawings

[0054] Figure 1 It is a flowchart of the perception data optimization method based on the autonomous UAV application of the present invention. Detailed Embodiment

[0055] The present invention will be further described below in conjunction with specific embodiments and the drawings.

[0056] According to the relevant introduction in the background section, the perception data optimization method provided by the disclosed embodiments of the present application, which is mainly applied to the optimization processing flow of the data output ports of the onboard software and hardware of the unmanned aerial vehicle (UAV), is mainly used to optimize the backend processing results of the data collected by the onboard sensors and solve the downstream response problems caused by the jump or interruption of the original data.

[0057] Embodiment 1

[0058] As Figure 1 shown, this embodiment provides a perception data optimization method based on autonomous UAV applications, including the following steps:

[0059] Step S100: Traverse the data according to the structure of the dynamic array; specifically, in this application, the common traversal methods such as forEach, map, filter, find, every, some, and reduce can be used to traverse the data according to the structure of the dynamic array.

[0060] Step S200: Set two arrays to track the data, where one array is the actual value array for recording the input data, and the other array is the modified value array for recording the data after correcting the newly incoming data; in this embodiment, by setting the actual value array and the modified value array, abnormal situations with relatively large offsets from the previous several data can be discarded according to the preset optimization steps, reducing the abrupt changes in the data.

[0061] The specific content of the modified value array recording the data after correcting the newly incoming data includes:

[0062] Step S210: Determine whether the degree of data offset of the newly incoming data is within the preset data offset range. If not, go to step S220; if so, go to step S230.

[0063] Step S220: Optimize the data with data offset, correct the data according to the degree of data offset, and record the corrected data in the modified value array.

[0064] Step S230: Record the newly incoming data in the modified value array.

[0065] It should be noted that during the process of the modified value array recording the data after correcting the newly incoming data, there are at least three problems. The first is how to determine whether the degree of data offset of the newly incoming data is within the preset data offset range. The second is how to optimize the data with data offset and correct the data according to the degree of data offset. The third is how to determine whether the data after correction of the newly incoming data is reasonable. The following mainly elaborates on the first two problems, and the third problem will be further explained in the subsequent embodiments and will not be specifically described here.

[0066] For Problem 1, in this embodiment, the confidence level of the reference input data is used as the basis for determining whether the degree of data deviation of the newly incoming data is within the preset data deviation range. For example, when the confidence level of the newly incoming data in the current frame is smaller or larger than the preset confidence level, it is determined at this time that it is necessary to enter step S220 to optimize the data with data deviation and correct the data according to the degree of data deviation; when the confidence level of the newly incoming data in the current frame is within the preset confidence level range, it can be determined that there is no need to optimize the newly incoming data. For example, there is no data input for the newly incoming data itself or the data change trend is already within a small deviation range. At this time, directly enter step S230 to record the newly incoming data into the modified value group.

[0067] For Problem 2, in step S220 of this embodiment, optimizing the data with data deviation and correcting the data according to the degree of data deviation, and recording the corrected data into the modified value group specifically includes:

[0068] If the deviation of the newly incoming data is large, compensate the data and add it to the modified value group;

[0069] Preferably, compensating the data may specifically include:

[0070] Step S211: Take the minimum second moment of the data of the previous several frames, solve the data of the current frame using this data and output it, and add the output result to the modified value group.

[0071] Exemplarily, the previous several frames of data may be 5 frames forward including the current new data. The minimum second moment is in the form of obtaining the central moment. First, obtain the first moment (expectation) of the previous 5 frames of data, and then obtain the second central moment (variance) based on the expectation. Among them, the second central moment represents the distribution dispersion degree of the original data (the previous 5 frames of data). It should be noted here that since the deviation of the newly incoming data is large, it is discarded during calculation, and the data is stored in the actual value group. At the same time, it is output in the form of the first moment expectation value ± the second central moment value in the modified value group.

[0072] In the above process, the confirmation of the ± sign is determined by the positive and negative deviation of the 5 frames of data based on the first moment, that is, if most of the data is greater than the expected value, the first moment expectation value + the second central moment value is output; if most of the data is less than the expected value, the first moment expectation value - the second central moment value is output; if a negative value appears at this time, the value is set to 0. This method of supplementing and outputting data using the foregoing results in the current frame discards the abnormal situations with relatively large deviations from the previous several data compared with the filter, reduces the abrupt changes, and has higher efficiency.

[0073] If the deviation of the newly incoming data is small, smooth the data and add it to the modified value group.

[0074] Preferably, smoothing the data may specifically include:

[0075] Step S212: Estimate the data using a state estimation method based on the filtering idea, replace the input value with the estimated value and output it, and add the output result to the modified value group. In the above steps, when estimating the data using a state estimation method based on the filtering idea, a Kalman filter can be specifically used. The input of the filter is the output data of the previous frame and the input data of the current frame. Taking an example, since the input data is usually convergent, that is, the difference between the input value groups is relatively small. Suppose there are two actual value results for the first frame and the second frame, and the error distributions of these two results converge to the inherent error. When the actual value of the third frame is input, use these results to predict the data of the third frame. First, perform a Kalman filter based on the predicted data and the actual data, and then use the result as the final value of the third frame data, that is, the modified value is output. Further, due to the existence of the inherent error of the input of the current frame, the initial error can be set to 0.25 when the filter performs filtering, so as to obtain a new fusion output. In the case of continuous data, the inherent error will gradually converge to a smaller value; in the case of non - continuous data, when the filter is enabled again, the inherent error will still be initialized to 0.25. In this embodiment, by using the filtering method, the data can be fused, making the data curve smoother, enabling a better transition, and avoiding the possible small fluctuations of the data and the sawtooth situation of the curve.

[0076] Step S213: Add the current gain result to the buffer. If the results are continuous, update it; if not, clear the result after a certain number of frames are satisfied. It should be noted that the gain result is the Kalman Gain calculated after the filter performs filtering. As mentioned above, since the data may become severely unstable, the input data of the filter may be either continuous or discrete and discontinuous; when the data is discrete and discontinuous, if the state estimation method based on the filtering idea is used to estimate the data through the method of step S212 at this time, the referenceability of the filtering result will become poor. Based on this, in this embodiment, through step S213, when the data is discrete and discontinuous, the gain is cleared and waiting for the next admission calculation, which further avoids the possible small fluctuations of the data and makes the data curve smoother.

[0077] It should be noted that since the confidence level of the input data is used as the basis for judging whether the degree of deviation of the newly incoming data is within the preset data deviation range in this embodiment, in a preferred embodiment, whether in the compensation layer logic algorithm with a large deviation or in the smoothing layer logic algorithm with a small deviation, before performing the corresponding algorithm steps, it may further include:

[0078] Judge whether the confidence level of the input data is small. If it is small, give a data warning.

[0079] In this preferred embodiment, the data warning refers to output in the form of printed information. For example, "The current data confidence level is low" is printed on the screen end, but it does not affect the subsequent logic. When such a label description appears in the original data, the algorithm processing will tend to modify the original data. Therefore, a warning needs to be given, and at the same time, the confidence level is given after the final result is calculated, and it is determined by subsequent algorithm processing or manual judgment whether it converges and is legal.

[0080] As mentioned above, since the algorithms for modifying the original data in both the compensation layer logic algorithm with a large offset and the smoothing layer logic algorithm with a small offset are evaluated or estimated based on the algorithms of the previous several frames. By judging in advance whether the confidence level of the input data is small, if it is small, a data warning is given, and it is possible to evaluate whether it is legal and credible to modify the subsequent frame data by means of the confidence level of each frame of data itself.

[0081] Step S300: Compare the data of the two arrays, and select the data of the corresponding frame numbers of different arrays for output according to the offset situation and the confidence level of the data of the corresponding frame numbers in the two arrays.

[0082] Specifically, it includes:

[0083] Step S310: Compare the data in the two arrays frame by frame to judge whether the values of the data of the corresponding frame numbers are the same;

[0084] Step S320: Judge the confidence level evaluation grade of the data of the current frame number in the actual value group;

[0085] Specifically, in step S321, establish the confidence level evaluation grade of the newly entered data. The confidence level evaluation grade of the data of the current frame number in the actual value group includes at least low, relatively low, relatively high, and high. It should be noted that the confidence level evaluation grade of the data here should establish a relationship with the confidence level of the current data. Specifically in this embodiment, when the confidence level of the data is in the range of 0 to 0.25, the default confidence level evaluation grade of the data is low; when the confidence level of the data is in the range of 0.25 to 0.5 (excluding 0.25), the default confidence level evaluation grade of the data is relatively low; when the confidence level of the data is in the range of 0.5 to 0.75 (excluding 0.5), the default confidence level evaluation grade of the data is medium; when the confidence level of the data is in the range of 0.75 to 1 (excluding 0.75 and 1), the default confidence level evaluation grade of the data is relatively high; when the confidence level of the data is 1, the default confidence level evaluation grade of the data is high.

[0086] Step S322: Judge the size of the confidence level of the newly entered data in the actual value group;

[0087] Step S323: Determine the confidence evaluation level of the newly entered data according to the size of the confidence of the newly entered data in the actual value group. For example, when the confidence of the newly entered data is 0.6, the confidence evaluation level of the current data should be medium;

[0088] In this embodiment, two dynamic arrays are set to track the data. One array only records the original data, and the other array records the modified data. For each frame of data in each array, the size of the data offset is evaluated by evaluating the confidence of the data. When the data offset is large, the compensation layer logic and algorithm with a large offset is used to correct the data, and the abnormal situations with a large offset relative to the previous several data are discarded, reducing the abrupt change of the data. When the data offset is small, the smoothing layer logic and algorithm with a small offset is used to modify the data, making the result more credible, thereby solving the downstream response problem caused by the jump or interruption of the original data collected during the high-speed flight of the UAV body.

[0089] Embodiment 2

[0090] It is basically the same as Embodiment 1. The difference is that in step S330 of this embodiment, selecting the data of the corresponding frame number in different arrays as the output data according to whether the values of the corresponding frame numbers are the same and the output confidence of different arrays specifically includes:

[0091] If the values of the current frame number data in the actual value group and the modified value group are the same, and the output confidence of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group as the output data;

[0092] If the values of the current frame number data in the actual value group and the modified value group are the same, and the output confidence of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the actual value group as the output data;

[0093] If the values of the current frame number data in the actual value group and the modified value group are different, further processing is performed;

[0094] First, determine the size of the difference value between the values of the current frame number data in the actual value group and the modified value group, where:

[0095] If the values of the current frame number data in the actual value group and the modified value group are different and the difference value is small, and the output confidence of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group as the output data;

[0096] If the values of the current frame number data in the actual value group and the modified value group are different and the difference value is large, and the output confidence of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the modified value group as the output data;

[0097] If the values of the current frame number data in the actual value group and the modified value group are different and the difference is large, and the output confidence level of the current frame number data in the actual value group is high or relatively high, select the data of the current frame number in the actual value group or the modified value group as the output data according to the situation again;

[0098] If the values of the current frame number data in the actual value group and the modified value group are different and the difference is large, and the output confidence level of the current frame number data in the actual value group is low or relatively low, select the data of the current frame number in the modified value group as the output data. Specifically, it includes:

[0099] Calculate the root mean square error RMSE for the data of the previous several frames;

[0100] If the root mean square error RMSE of the data of the previous several frames is less than or equal to the preset root mean square error, select the data of the current frame number in the actual value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to a relatively high level, and delete the data of the current frame number in the modified value group;

[0101] If the root mean square error RMSE of the data of the previous several frames is greater than the preset root mean square error, judge the confidence level of the data of the previous several frames; among them:

[0102] If the confidence levels of the data of the previous several frames are all high or relatively high, select the data of the current frame number in the actual value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to medium; otherwise:

[0103] Select the data of the current frame number in the modified value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to a relatively low level.

[0104] It should be noted that in order to solve the aforementioned problem of how to judge whether the corrected data is reasonable after correcting the newly incoming data, for the situation where the values of the corresponding frame numbers in the actual value group and the modified value group are different and the difference is large, since there is an original confidence level information in the actual value group, and the confidence level of the modified value group will be output in the aforementioned algorithm, the relationship between the actual value group and the modified value group utilizes the principle and idea of fuzzy mathematics;

[0105] If the confidence level of the actual value group is high but the gap from the modified value is large, then take the results and confidence level of the first 5 frames of data, calculate the root mean square error RMSE of the data results of these 5 frames. If the RMSE value is less than 0.25, it is considered that the current data is valid, and the data of the actual value group is still used, and the corresponding modified value is discarded, and the confidence evaluation level is output as the higher or high grade (that is, greater than 0.75); if the value is greater than 0.25, then examine its confidence level. If the confidence levels of all 5 frames are in the high or relatively high range, the data of the actual value group is also output, and the confidence evaluation level is output as medium; if there are medium, low, or relatively low cases in the confidence evaluation levels of these 5 frames, at this time, the confidence level is output as relatively low, and the value is the data corresponding to the modified value group.

[0106] If the confidence level of the actual value group is low and the gap from the modified value is large, in this case, the data in the actual value group cannot be trusted. Therefore, without considering different situations, the output is directly the modified value, and the confidence evaluation level is the confidence level of the final output data of the previous frame. This is because when the data cannot be confirmed as credible in this frame, the value is no longer the key, and the same confidence level as the previous frame is adopted here as a transition.

[0107] In this embodiment, according to whether the values corresponding to the frame numbers are the same and different situations of the output confidence levels of different arrays, the data of the corresponding frame numbers of different arrays are selected as the output data. At the same time, with the help of the principle and idea of fuzzy mathematics, the problem of whether it is reasonable in the subsequent frame optimization step after the newly entered data is corrected and the corrected data is used is mainly considered, so that the data results output by the perception data optimization method based on autonomous unmanned aerial vehicle applications in this embodiment are more reliable, and the spatio-temporal efficiency and accuracy are higher.

[0108] The examples described in the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design idea of the present invention, various deformations and improvements made by those skilled in the art to the technical solutions of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for optimizing perception data based on the application of autonomous drones, characterized in that: It includes the following steps: Step S100: Traverse the data according to the structure of the dynamic array; Step S200: Set two arrays to track the data. One array is the actual value array for recording newly entered data, and the other array is the modified value array for recording the data after correcting the newly entered data; Step S300: Compare the data of the two arrays, and select the data corresponding to the frame numbers of different arrays for output according to the offset situation and confidence level of the data corresponding to the frame numbers in the two arrays; In step S200, the modified value array records the data after correcting the newly entered data, which specifically includes: Step S210: Judge whether the degree of offset of the newly entered data is within the preset data offset range. If not, enter step S220; if so, enter step S230; Step S220: Optimize the data with data offset, correct the data according to the degree of data offset, and record the corrected data in the modified value array; Step S230: Record the newly entered data in the modified value array.

2. The perception data optimization method based on autonomous drone applications according to claim 1, wherein In step S220, optimizing the data with data offset, correcting the data according to the degree of data offset, and recording the corrected data in the modified value array specifically includes: If the offset of the newly entered data is large, compensate the data and add it to the modified value array; If the offset of the newly entered data is small, smooth the data and add it to the modified value array.

3. The perception data optimization method based on autonomous drone applications according to claim 2, characterized in that: If the offset of the newly entered data is large, directly compensating the result and adding it to the modified value array specifically includes: Step S211: Take the minimum second moment of the data of the previous several frames, solve the data of the current frame using this data and output it, and add the output result to the modified value array.

4. The perception data optimization method based on the application of autonomous drones according to claim 2, wherein: If the offset of the newly entered data is small, smoothing the data and adding it to the modified value array specifically includes: Step S212: Estimate the data using the state estimation method with the idea of filtering, use the estimated value to replace the input value and output it, and add the output result to the modified value array; Step S213: Add the current gain result to the buffer, update it if the results are continuous, and clear the result after a certain number of frames if they are not continuous.

5. The perception data optimization method based on the application of autonomous drones according to claim 1, characterized in that: In step S300, comparing the data of the two arrays, and selecting the data corresponding to the frame numbers of different arrays for output according to the offset situation and confidence level of the data corresponding to the frame numbers in the two arrays specifically includes: Step S310: Compare the data in the two arrays frame by frame to judge whether the values of the data corresponding to the frame numbers are the same; Step S320: Judge the confidence evaluation level of the data of the current frame number in the actual value array; Step S330: Select the data corresponding to the frame numbers of different arrays as the output data according to whether the values corresponding to the frame numbers are the same and the output confidence levels of different arrays.

6. The perception data optimization method based on autonomous drone applications according to claim 5, wherein In step S320, judging the confidence evaluation level of the data of the current frame number in the actual value array specifically includes: Step S321: Establish the confidence evaluation level of the newly entered data. The confidence evaluation level of the data of the current frame number in the actual value array includes at least low, relatively low, medium, relatively high, and high; Step S322: Judge the size of the confidence of the newly entered data in the actual value array; Step S323: Determine the confidence evaluation level of the newly entered data according to the magnitude of the confidence of the newly entered data in the actual value group.

7. The perception data optimization method based on the application of autonomous drones according to claim 5, characterized in that, In step S330, selecting the data of the corresponding frame number in different arrays as the output data according to whether the values of the corresponding frame numbers are the same and the output confidence of different arrays specifically includes: If the value of the current frame number data in the actual value group is the same as that in the modified value group, and the output confidence of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group as the output data; If the value of the current frame number data in the actual value group is the same as that in the modified value group, and the output confidence of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the actual value group as the output data.

8. The perception data optimization method based on the application of autonomous drones according to claim 5, characterized in that In step S330, selecting the data of the corresponding frame number in different arrays as the output data according to whether the values of the corresponding frame numbers are the same and the output confidence of different arrays also includes: If the value of the current frame number data in the actual value group is different from that in the modified value group, judge the magnitude of the difference value between the values of the current frame number data in the actual value group and the modified value group, where: If the value of the current frame number data in the actual value group is different from that in the modified value group and the difference value is small, and at the same time the output confidence of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group as the output data; If the value of the current frame number data in the actual value group is different from that in the modified value group and the difference value is large, and at the same time the output confidence of the current frame number data in the actual value group is high or at a relatively high level, select the data of the current frame number in the actual value group or the modified value group as the output data according to the situation; If the value of the current frame number data in the actual value group is different from that in the modified value group and the difference value is large, and at the same time the output confidence of the current frame number data in the actual value group is low or at a relatively low level, select the data of the current frame number in the modified value group as the output data.

9. The perception data optimization method based on autonomous drone applications according to claim 8, wherein If the value of the current frame number data in the actual value group is different from that in the modified value group and the difference value is large, and at the same time the output confidence of the current frame number data in the actual value group is high or at a relatively high level, selecting the data of the current frame number in the actual value group or the modified value group as the output data according to the situation specifically includes: Calculate the root mean square error RMSE for the data of the previous several frames; If the root mean square error RMSE of the data of the previous several frames is less than or equal to the preset root mean square error, select the data of the current frame number in the actual value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to a higher level, and delete the data of the current frame number in the modified value group; If the root mean square error RMSE of the data of the previous several frames is greater than the preset root mean square error, judge the confidence level of the data of the previous several frames; where: If the confidence levels of the data of the previous several frames are all high or at a relatively high level, select the data of the current frame number in the actual value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to medium; otherwise: Select the data of the current frame number in the modified value group as the output data, and at the same time adjust the confidence level of the current frame data in the actual value group to a relatively low level.

Images