CN113076860A - Bird detection system under field scene - Google Patents

Bird detection system under field scene Download PDF

Info

Publication number
CN113076860A
CN113076860A CN202110344311.XA CN202110344311A CN113076860A CN 113076860 A CN113076860 A CN 113076860A CN 202110344311 A CN202110344311 A CN 202110344311A CN 113076860 A CN113076860 A CN 113076860A
Authority
CN
China
Prior art keywords
bird
data
detection
image
data set
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110344311.XA
Other languages
Chinese (zh)
Other versions
CN113076860B (en
Inventor
腊孟珂
肖伟康
钟稚昉
杨仕雄
王言
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University Environmental Planning And Design Institute Group Co Ltd
Original Assignee
Nanjing University Environmental Planning And Design Institute Group Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University Environmental Planning And Design Institute Group Co Ltd filed Critical Nanjing University Environmental Planning And Design Institute Group Co Ltd
Priority to CN202110344311.XA priority Critical patent/CN113076860B/en
Publication of CN113076860A publication Critical patent/CN113076860A/en
Application granted granted Critical
Publication of CN113076860B publication Critical patent/CN113076860B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a detection method of a bird detection system in a field scene, which comprises the following steps: s1, collecting video data in an actual scene and bird pictures and video data existing on a network; s2, carrying out data preprocessing on the acquired bird data by using data preprocessing software; s3, performing data enhancement on the processed bird picture data set; s4, training a deep learning target detection model by using the bird data set after data enhancement and the background data set without birds; s5, deploying the trained target detection model on edge equipment in an actual field environment; s6, acquiring video data of the bird protection area in a field scene by using a camera; and S7, detecting the target according to the image by using the deep learning target detection model, and returning the detection result to the central server. The invention reduces the cost of manual detection, improves the detection effect of the model, obtains the result in accordance with the actual scene, and reduces the processing pressure of the central server.

Description

Bird detection system under field scene
Technical Field
The invention belongs to the technical field of bird detection methods, and particularly relates to a bird detection system in a field scene.
Background
The discovery and statistics of birds in protected areas has been an important aspect of environmental protection. By recording the number of the bird groups, the bird occurrence time and the residence time, the current health state of the bird groups and the information such as the ecological environment quality of the protected area can be effectively reflected, and the professional can conveniently develop the environment protection work. However, in the conventional method, video data shot by a camera in an environment protection area is classified and counted manually, and the method needs experts to spend a lot of time on processing the video data, and is time-consuming, labor-consuming and large in workload. Therefore, the birds in the protected area can be automatically detected by adopting a target detection algorithm based on deep learning in the field of artificial intelligence.
Target detection refers to finding out an object of interest in an image, determining the category and position of the object, for example, detecting whether birds exist in the image and marking the position information of the birds in the image. Most of traditional target detection algorithms are designed manually to process images, only shallow features can be extracted, detection accuracy is low, various noise interferences can be caused in actual environments, and generalization capability is poor. The deep learning model has unusual expression in the aspect of extracting high-level features and representing capability of complex features, and more target detection algorithms based on deep learning are developed in recent years. The target detection algorithm based on deep learning can automatically extract features, has stronger generalization capability, and is more suitable for deployment and application in actual environment.
The target detection task focuses on the accuracy rate and the recognition speed of detection. Due to the influences of the resolution of the detected image, the size of the target and the complexity of the environment, great difficulty is brought to the target detection. The earliest two stages of generating a candidate window and classifying objects in the candidate window are divided into two stages by the two stages of the initial two-stage target detection algorithm R-CNN, compared with the traditional method, the identification accuracy is obviously improved, and the detection accuracy and the detection speed are improved to a certain extent by a series of two-stage algorithms subsequently proposed based on the R-CNN. However, due to the self-limitation, the calculation cost of the two-stage algorithm is too high, the detection time of one image is too long, the practical application is not met, and meanwhile, the relative position information of the local target in the whole body cannot be utilized. Researchers have therefore proposed a one-stage target detection algorithm, which is represented as a YOLO series, and such algorithms place the whole process in a network for processing, directly obtain candidate regions and target categories, and greatly improve the detection speed.
Considering a series of characteristics of complex environment, small target bird size, limited calculation capacity of edge equipment and the like in an environment protection area, it is necessary to design a target detection system under a field scene according to the existing target detection algorithm and an actual scene.
Disclosure of Invention
The invention aims to overcome partial defects of the application of the current target detection method based on deep learning under the field condition.
In order to solve the defects, the invention discloses a bird detection system under a field scene, which can be practically deployed and applied by combining with an One-Stage target detection model.
In order to achieve the purpose, the invention adopts the following technical scheme:
a bird detection system under a field scene, wherein a detection method of the bird detection system comprises the following steps:
s1, acquiring bird video data and bird picture data; including the bird data D shot by a camera and existing unmarked in the InternetunlabeledbirdAnd acquiring a marked bird picture data set D from the Internetlabeled1Acquiring background video data D from a camera shot without birdsvideoenv；
S2, collecting the unlabelled bird data DunlabeledbirdUsing the written data preprocessing software to carry out data preprocessing to obtain a bird picture data set Dlabeled2(ii) a Butt-miningCollected background video data D not including birdsvideoenvPreprocessing software written by data is used for preprocessing the data to obtain a background picture data set D not containing birdsnegative；
S3, marking bird picture data set Dlabeled1And Dlabeled2Carrying out data enhancement to obtain an expanded bird data set Dpositive；
S4 bird data set D enhanced by datapositiveAnd a background data set D not containing birdsnegativeModel for training deep learning target detection Modelbird；
S5, detecting the Model of the trained targetbirdThe method comprises the following steps of deploying on edge equipment in an actual field environment;
s6, acquiring video data of the bird protection area in a field scene by using a camera;
s7, Model for detecting target by using deep learning targetbirdAnd detecting the target according to the image, and returning the detection result to the central server.
In order to optimize the technical scheme, the specific measures adopted further comprise:
further, step S1 includes:
s11, the camera is deployed in the actual environment protection area;
s12, shooting a bird group and an environment background in the protected area by a camera;
s13, acquiring marked bird images, unmarked bird images and video data from public data sets in the Internet;
s14, forming a data set D by the marked bird picture data disclosed in the Internetlabeled1。
S15, the bird video data shot by the camera and the unmarked bird image and video data collected from the network form a data set Dunlabeledbird；
S16, background video data containing no bird shot by camera is formed into data set Dnegative。
Further, the data preprocessing process of step S2 includes:
designing and using data preprocessing software to frame background video data not containing birds and non-labeled bird video data, and respectively storing the frame background video data not containing birds and the image data containing birds, wherein the background image data not containing birds forms a data set Dnegative；
Performing quality detection on all unmarked image data containing birds, and removing images with too low bird target pixels;
labeling the image data after quality detection, recording bird position bounding-box information, and obtaining a labeled bird picture data set Dlabeled2。
Further, step S2 further includes:
writing video data preprocessing software;
performing image extraction on the video data by using video data preprocessing software;
and (3) using video data preprocessing software to carry out quality detection on the image data and remove the images with too low bird target pixels.
And using video data preprocessing software to perform data marking on the image data and record bird position information.
Further, the data enhancement process of step S3 includes:
performing data enhancement on the data set by using an image cutting mode, wherein the mode is to generate a rectangular frame with a smaller size than the image, randomly cutting the image, and finally taking the image in the rectangular frame as training data;
performing data enhancement on the data set by using an image turning mode, wherein the image is turned left and right, and the images before and after turning are simultaneously used as training data;
the data set is enhanced by using an image whitening mode, namely, the image is whitened, namely, the image is normalized into Gaussian (0, 1) distribution, and the images before and after normalization are simultaneously used as training data.
Further, step S4 includesData set D enhanced with datapositiveAnd a background data set DnegativeTraining a target detection network Yolov4, comprising the following steps:
s41, data set DpositiveAnd DnegativeMixed composition training set Dtrain；
S42, data set DtrainDividing the training set, the verification set and the test set;
s43, modifying the YOLOv4 network into a two-classification network;
s44, according to the data set DpositiveCalculating and setting an anchor frame value as a Yolov4 anchor frame value;
s45, setting parameters of the target detection network;
s46, training the target detection network by using the training set and the verification set;
s47, verifying the target detection effect by using the test set;
s48, repeating the steps S45-S47 until the proper parameters are searched.
Further, in step S44, calculating the anchor block value using the kmeans method includes:
for data set DpositiveCounting all bounding box information in the database, and dividing the bounding box information into 9 categories by using a kmeans algorithm; the average bounding box size among the 9 classes is calculated as the corresponding anchor box value size.
Furthermore, the edge device is a gateway and comprises a CPU and a GPU hardware;
the edge device directly controls the camera;
the edge device can transmit the result obtained by the deep learning target detection model back to the central server.
Further, the camera is deployed in a bird protection area of a field scene;
performing video plug flow by using ffmpeg and rtsp protocols;
images in the video stream are acquired using opencv.
Further, in step S7, a deep learning object detection Model is usedbirdThe detection of the target by image is specifically as follows:
sampling images acquired by a field camera, and reducing the calculation pressure of edge equipment;
using a trained target detection Model for the imagebirdDetecting to obtain a detection result;
the detection result comprises whether the image has the bird or not, and position information and confidence coefficient of the bird;
denoising the detection result to filter out the detection result with noise information and low confidence coefficient in the actual scene;
locally storing the image of the bird detected in the result, and storing the position information of the bird in the image;
and feeding back the locally stored image containing the birds and the corresponding position information to the central server.
Further, the denoising process includes:
model using target detection ModelbirdDetecting the test set to obtain a detection result;
carrying out statistical analysis on the detection results, and classifying according to the size of the detection frame;
standardizing the confidence degrees of the detection frames in the same size range to form normal distribution;
taking the normal distribution as a probability density function, and calculating the corresponding lowest confidence coefficient when the probability is 0.95;
taking the obtained lowest confidence as a confidence threshold of the detection frame in the size range;
and finding a corresponding threshold value according to the size of the detection frame by the actual detection result, returning the detection result with the confidence coefficient larger than the threshold value as a final result, and discarding the detection result with the confidence coefficient smaller than the threshold value.
The invention has the beneficial effects that: the invention uses the camera to monitor the bird protection area, and uses the target detection method to detect the birds appearing in the video, thereby reducing the cost of manual detection; data training in an actual scene is used, and preprocessing and data enhancement operations are performed on the data, so that the detection effect of the model is improved; denoising the detection result to obtain a result in accordance with an actual scene; and the target detection model is deployed on the edge equipment, so that the processing pressure of the central server is relieved.
Drawings
FIG. 1 is a flow chart of the operation of the system of the present invention;
FIG. 2 is a schematic diagram of data preprocessing software programmed in the present invention;
FIG. 3 is a diagram showing the labeling effect of the data preprocessing software written in the present invention;
FIG. 4 is a representation of a portion of a sample in an avian target detection dataset as used in the present invention;
FIG. 5 is a schematic diagram of a training process of the target detection model according to the present invention;
FIG. 6 is a schematic diagram of a process for detecting an object on the edge device according to the present invention;
fig. 7 is a detection effect display diagram of bird target detection used in the present invention.
Detailed Description
The present invention will now be described in further detail with reference to the accompanying drawings.
It should be noted that the terms "upper", "lower", "left", "right", "front", "back", etc. used in the present invention are for clarity of description only, and are not intended to limit the scope of the present invention, and the relative relationship between the terms and the terms is not limited by the technical contents of the essential changes.
Fig. 1 is a flow chart of a bird detection system in a field scene according to the present invention, which includes the following 11 steps.
Step 1, collecting data, and constructing a bird data set and a background data set for subsequent data processing. The bird data collected in the example comprise three sources, one of the three sources is that a public picture data set on the internet is collected manually, the public picture data set comprises data sets such as COCO and VOC, and the marked bird data in the data set is taken out to form Dlabeled1(ii) a Secondly, the bird video data are published on the internet; thirdly, the video camera is fixed in the area where birds frequently appear in the environment protection area from the bird video data shot by the video camera in the environment protection areaThe domain increases the probability of birds appearing in the video, and the latter two data sets jointly form a data set Dunlabeledbird. The background data collected in this example are from the environmental video data of birds not included taken by the cameras in the environmental protection area, constituting a data set Dvideoenv
Step 2, data preprocessing and background video data set D processingvideoenvObtaining a background image data set Dnegative. Processing a data set DunlabeledbirdA noise-free and directly usable data set is obtained for subsequent data enhancement. Data pre-processing in this example includes video data framing, data quality detection, and data annotation. The method specifically comprises the steps of writing data preprocessing software R shown in figure 2, opening background video data by using the R, storing a current frame image, and finally obtaining a background picture data set Dnegative. Opening bird video data by using the R, detecting according to frames, and directly removing birds if one frame of image does not contain birds; if the frame contains birds but the pixels are low, the frame is also rejected; if a bird is present in the frame and the lowest pixel size is satisfied, the position of the bird is marked on the image with a mouse using a marking function, as shown in FIG. 3. The next frame is then processed until a segment of video processing is complete, all images containing birds and corresponding bird location information are saved locally. Using R to open unlabeled bird image data, the standard approach is similar to video data. Finally obtaining a labeled bird image data set Dlabeled2。
And 3, data enhancement, namely performing data enhancement on the preprocessed data set to increase the size of the data set and enlarge the posture of the target displayed in the image. In this example, the operations of cutting, turning over and whitening are respectively performed on the image; respectively cutting the images in a mode of generating a rectangular frame with a size smaller than that of the images, randomly cutting the images, and finally taking the images in the rectangular frame as training data; turning the image in a mode of turning the image left and right, and simultaneously using the image before and after turning as training data; the whitening operation is applied to the image in such a way as to normalize the image itselfAnd forming Gaussian (0, 1) distribution, and simultaneously using the images before and after normalization as training data. Finally, a marked image data set D with 16981 birds is obtainedpositiveAs shown in fig. 4.
Step 4, training a deep learning target detection model, which is embodied in the embodiment that bird picture data and background picture data are used as positive and negative samples to be mixed to obtain a complete data set DtrainAnd randomly dividing the training set into training sets according to the ratio of 3: 1: and (4) verification set: and (5) testing the set. Wherein, the training set and the verification set are used as the input of the model to complete the training process of the model; the test set is used for evaluating the model and judging whether the model achieves the required effect. The model is selected as YOLOV4, and the training process of the model is shown in fig. 5, and includes the following steps:
and 4-1, modifying a classification layer in the network to enable the classification layer to perform a classification task.
Step 4-2, setting model parameters, specifically setting a learning rate of 0.001, a batch size of 64, a weight attenuation of 0.0005, training steps of 2000, and selecting DarknetNet53 for the backbone network;
step 4-3, according to bird picture data set DpositiveThe bounding box information is divided into 9 classes by using a kmeans clustering method, the mean value of all the bounding box information in each class is used as the final anchor frame value, and the obtained 9 anchor frame values are used as the prior knowledge of the target detection model to replace the default anchor frame value.
Step 4-4, inputting the data of the training set and the verification set into the model;
step 4-5, training YOLOV4 on an Nvidia 2080ti display card by using an open source pre-training model, and storing a weight file obtained by training after the training is finished;
4-6, verifying the effect of the model after training by using the weight file and the test set obtained by training;
step 4-7, judging a test result, and if the Average Precision of the mAP (mean Average Precision) reaches 0.7, regarding the parameters and the model obtained by training at the moment as the optimal result; otherwise, modifying the parameter value and repeating the steps 4-4 to 4-7.
And 5, deploying the algorithm in the actual field environment, specifically in the example, deploying the algorithm on edge equipment, deploying the edge equipment into a chassis existing in an environment protection area, wherein the edge equipment is a gateway and can directly control the camera and acquire the video in the actual environment from the camera. The gateway device comprises a CPU10700K and a GPU GeForce RTX 2080ti, and is used for providing computational power requirements of a target detection network, and all detection processes are completed on the edge device.
And 6, acquiring video data shot by the camera, wherein the camera is a dome camera in the embodiment, and can rotate and zoom to complete the omnibearing detection of the area near the dome camera. Video stream data acquired by the dome camera is transmitted through an rtsp protocol and is pushed to the network by ffmpeg, and then the edge device captures the video data from the rtsp stream through opencv and takes out images in the video data.
Step 7, processing the image using the target detection model, which is embodied as shown in fig. 6 in this example, includes the following steps:
step 7-1, regarding the image acquired by opencv, considering that the calculation power of edge equipment is limited, the algorithm detection speed needs to be improved through image down-sampling, in this example, the image is down-sampled to 576 x 576, and then the down-sampled image is input into the target detection algorithm;
7-2, detecting the down-sampled image by using the trained weight parameters stored by the edge end by using a target detection algorithm to obtain a corresponding detection result, wherein the detection result comprises a detection frame and the probability that birds are contained in the detection frame, judging the detection result, and detecting the next frame of image if the birds are not detected in the frame; if birds are detected in the current frame, noise possibly existing in an actual scene is considered, denoising processing needs to be carried out on a detection result, and the specific expression is that statistical analysis is carried out on collected test set data in advance, classification is carried out according to the size of detection frames in the detection result, and the confidence coefficients of the detection frames in the same size range are standardized to form normal distribution. And then, taking the normal distribution as a probability density function, calculating the corresponding lowest confidence coefficient when the probability is 0.95, and taking the obtained lowest confidence coefficient as a confidence coefficient threshold value of the detection frame in the size range. The actual detection result finds the corresponding threshold according to the size of the detection frame, the detection result with the confidence degree greater than the threshold is returned as the final result, the detection result with the confidence degree less than the threshold is discarded, and finally the frame and the position information of the corresponding bird are stored locally, which is specifically shown in fig. 7. The locally stored detection results are transmitted to the central server. The results, embodied in this example as locally stored, are the images that have been algorithmically detected as birds and de-noised, and the edge devices transmit these results over the network to the central server, helping the central server to count the data.
The method and the device monitor the protected area by using the camera, detect birds appearing in the video by using the target detection method, reduce the cost of manual detection, use data training in an actual scene, perform preprocessing and data enhancement operation on the data, and improve the detection effect of the model; denoising the detection result to obtain a result in accordance with an actual scene; and the target detection model is deployed on the edge equipment, so that the processing pressure of the central server is relieved.
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may be made by those skilled in the art without departing from the principle of the invention.

Claims (10)

1. A bird detection system under a field scene is characterized in that a detection method of the bird detection system comprises the following steps:
s1, acquiring bird video data and bird picture data; including unlabeled bird data DunlabeledbirdAcquiring a data set D from a publicly labeled bird picturelabeled1Collecting bags from camera shotsBackground video data D containing birdsvideoenv；
S2, collecting the unlabelled bird data DunlabeledbirdUsing data preprocessing software to carry out data preprocessing to obtain a bird picture data set Dlabeled2(ii) a For the collected background video data D not containing birdsvideoenvUsing data preprocessing software to carry out data preprocessing to obtain a background picture data set D not containing birdsnegative；
S3, marking bird picture data set Dlabeled1And Dlabeled2Carrying out data enhancement to obtain an expanded bird data set Dpositive；
S4 bird data set D enhanced by datapositiveAnd a background data set D not containing birdsnegativeModel for training deep learning target detection Modelbird；
S5, detecting the Model of the trained targetbirdThe method comprises the following steps of deploying on edge equipment in an actual field environment;
s6, acquiring video data of the bird protection area in a field scene by using a camera;
s7, Model for detecting target by using deep learning targetbirdAnd detecting the target according to the image, and returning the detection result to the central server.
2. The bird detection system of claim 1, wherein the bird data D not labeled in step S1unlabeledbirdIncluding unlabelled bird image data and video data obtained from the internet, and video data containing birds shot by a camera.
3. The bird detection system of claim 2 wherein the data preprocessing process of step S2 includes:
designing and using data preprocessing software to frame background video data not containing birds and unlabelled bird video data, and respectively storing the background video data not containing birds and the image data containing birds, wherein the background video data not containing birds and the image data containing birds are not containing birdsScene image data composition data set Dnegative；
Performing quality detection on all unmarked image data containing birds, and removing images with too low bird target pixels;
labeling the image data after quality detection, recording bird position bounding-box information, and obtaining a labeled bird picture data set Dlabeled2。
4. The bird detection system of claim 1 wherein the data enhancement of step S3 includes:
performing data enhancement on the data set by using an image cutting mode, wherein the mode is to generate a rectangular frame with a smaller size than the image, randomly cutting the image, and finally taking the image in the rectangular frame as training data;
performing data enhancement on the data set by using an image turning mode, wherein the image is turned left and right, and the images before and after turning are simultaneously used as training data;
the data set is enhanced by using an image whitening mode, namely, the image is whitened, namely, the image is normalized into Gaussian (0, 1) distribution, and the images before and after normalization are simultaneously used as training data.
5. The bird detection system of claim 1, wherein step S4 includes using data enhanced data set DpositiveAnd a background data set DnegativeTraining a target detection network Yolov4, comprising the following steps:
s41, data set DpositiveAnd DnegativeMixed composition training set Dtrain；
S42, data set DtrainDividing the training set, the verification set and the test set;
s43, modifying the YOLOv4 network into a two-classification network;
s44, according to the data set DpositiveCalculating and setting an anchor frame value as a Yolov4 anchor frame value;
s45, setting parameters of the target detection network;
s46, training the target detection network by using the training set and the verification set;
s47, verifying the target detection effect by using the test set;
s48, repeating the steps S45-S47 until the proper parameters are searched.
6. The bird detection system of claim 4 wherein in step S44, calculating an anchor frame value using the kmeans method includes:
for data set DpositiveCounting all the bounding box information in the database, and dividing the bounding box information into 9 categories by using a kmeans algorithm; the average bounding box size among the 9 classes is calculated as the corresponding anchor box value size.
7. The bird detection system of claim 1, wherein the edge device is a gateway comprising a piece of CPU and a piece of GPU hardware;
the edge device directly controls the camera;
the edge device can transmit the result obtained by the deep learning target detection model back to the central server.
8. The bird detection system of claim 1, wherein the camera is deployed in a bird protection area of a field scene;
performing video plug flow by using ffmpeg and rtsp protocols;
images in the video stream are acquired using opencv.
9. The bird detection system of claim 4, wherein in step S7, a deep learning target detection Model is usedbirdThe detection of the target by image is specifically as follows:
sampling images acquired by a field camera, and reducing the calculation pressure of edge equipment;
using a trained target detection Model for the imagebirdDetecting to obtainDetecting the detection result;
the detection result comprises whether the image has the bird or not, and position information and confidence coefficient of the bird;
denoising the detection result, and filtering out the detection result with noise information and low confidence coefficient in the actual scene;
locally storing the image of the bird detected in the result, and storing the position information of the bird in the image;
and feeding back the locally stored image containing the birds and the corresponding position information to the central server.
10. The bird detection system of claim 8, wherein the de-noising process comprises:
model using target detection ModelbirdDetecting the test set to obtain a detection result;
carrying out statistical analysis on the detection results, and classifying according to the size of the detection frame;
standardizing the confidence degrees of the detection frames in the same size range to form normal distribution;
taking the normal distribution as a probability density function, and calculating the corresponding lowest confidence coefficient when the probability is 0.95;
taking the obtained lowest confidence as a confidence threshold of the detection frame in the size range;
and finding a corresponding threshold value according to the size of the detection frame by the actual detection result, returning the detection result with the confidence coefficient larger than the threshold value as a final result, and discarding the detection result with the confidence coefficient smaller than the threshold value.
CN202110344311.XA 2021-03-30 2021-03-30 Bird detection system under field scene Active CN113076860B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110344311.XA CN113076860B (en) 2021-03-30 2021-03-30 Bird detection system under field scene

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110344311.XA CN113076860B (en) 2021-03-30 2021-03-30 Bird detection system under field scene

Publications (2)

Publication Number Publication Date
CN113076860A true CN113076860A (en) 2021-07-06
CN113076860B CN113076860B (en) 2022-02-25

Family

ID=76611769

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110344311.XA Active CN113076860B (en) 2021-03-30 2021-03-30 Bird detection system under field scene

Country Status (1)

Country Link
CN (1) CN113076860B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113486866A (en) * 2021-09-06 2021-10-08 南京天朗防务科技有限公司 Visual analysis method and system for airport bird identification
CN113780063A (en) * 2021-07-27 2021-12-10 深圳泰豪信息技术有限公司 Photovoltaic operation and maintenance control method based on video intelligent analysis

Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108509939A (en) * 2018-04-18 2018-09-07 北京大学深圳研究生院 A kind of birds recognition methods based on deep learning
CN108920855A (en) * 2018-07-12 2018-11-30 西安英特迈思信息科技有限公司 A kind of UAV Intelligent monitoring threshold based on probability density distribution determines method
CN109115770A (en) * 2018-06-14 2019-01-01 中科禾信遥感科技（苏州）有限公司 A kind of a wide range of crops remote-sensing monitoring method and device
CN109241897A (en) * 2018-08-29 2019-01-18 深圳华远云联数据科技有限公司 Processing method, device, gateway and the storage medium of monitoring image
CN109376605A (en) * 2018-09-26 2019-02-22 福州大学 A kind of electric inspection process image bird-resistant fault detection method
CN109615171A (en) * 2018-11-09 2019-04-12 阿里巴巴集团控股有限公司 Characteristic threshold value determines that method and device, problem objects determine method and device
CN109886083A (en) * 2019-01-03 2019-06-14 杭州电子科技大学 A kind of small face detecting method of real-time scene based on deep learning
CN110379143A (en) * 2019-07-12 2019-10-25 中铁工程机械研究设计院有限公司 A kind of alarm system Threshold for mechanical equipment
CN110378420A (en) * 2019-07-19 2019-10-25 Oppo广东移动通信有限公司 A kind of image detecting method, device and computer readable storage medium
CN110399868A (en) * 2018-04-19 2019-11-01 北京大学深圳研究生院 A kind of seashore wetland birds detection method
CN110889841A (en) * 2019-11-28 2020-03-17 江苏电力信息技术有限公司 YOLOv 3-based bird detection algorithm for power transmission line
CN111709421A (en) * 2020-06-18 2020-09-25 深圳市赛为智能股份有限公司 Bird identification method and device, computer equipment and storage medium
CN111709374A (en) * 2020-06-18 2020-09-25 深圳市赛为智能股份有限公司 Bird condition detection method and device, computer equipment and storage medium
CN111753702A (en) * 2020-06-18 2020-10-09 上海高德威智能交通系统有限公司 Target detection method, device and equipment
CN111768246A (en) * 2020-06-30 2020-10-13 创新奇智(北京)科技有限公司 Data processing method, model establishing device and electronic equipment
CN112183191A (en) * 2020-08-18 2021-01-05 天津大学 Method for detecting bird nest in power transmission line based on countermeasure training
CN112395959A (en) * 2020-10-30 2021-02-23 天合云能源互联网技术(杭州)有限公司 Power transformer fault prediction and diagnosis method and system based on audio features

Patent Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108509939A (en) * 2018-04-18 2018-09-07 北京大学深圳研究生院 A kind of birds recognition methods based on deep learning
CN110399868A (en) * 2018-04-19 2019-11-01 北京大学深圳研究生院 A kind of seashore wetland birds detection method
CN109115770A (en) * 2018-06-14 2019-01-01 中科禾信遥感科技（苏州）有限公司 A kind of a wide range of crops remote-sensing monitoring method and device
CN108920855A (en) * 2018-07-12 2018-11-30 西安英特迈思信息科技有限公司 A kind of UAV Intelligent monitoring threshold based on probability density distribution determines method
CN109241897A (en) * 2018-08-29 2019-01-18 深圳华远云联数据科技有限公司 Processing method, device, gateway and the storage medium of monitoring image
CN109376605A (en) * 2018-09-26 2019-02-22 福州大学 A kind of electric inspection process image bird-resistant fault detection method
CN109615171A (en) * 2018-11-09 2019-04-12 阿里巴巴集团控股有限公司 Characteristic threshold value determines that method and device, problem objects determine method and device
CN109886083A (en) * 2019-01-03 2019-06-14 杭州电子科技大学 A kind of small face detecting method of real-time scene based on deep learning
CN110379143A (en) * 2019-07-12 2019-10-25 中铁工程机械研究设计院有限公司 A kind of alarm system Threshold for mechanical equipment
CN110378420A (en) * 2019-07-19 2019-10-25 Oppo广东移动通信有限公司 A kind of image detecting method, device and computer readable storage medium
CN110889841A (en) * 2019-11-28 2020-03-17 江苏电力信息技术有限公司 YOLOv 3-based bird detection algorithm for power transmission line
CN111709421A (en) * 2020-06-18 2020-09-25 深圳市赛为智能股份有限公司 Bird identification method and device, computer equipment and storage medium
CN111709374A (en) * 2020-06-18 2020-09-25 深圳市赛为智能股份有限公司 Bird condition detection method and device, computer equipment and storage medium
CN111753702A (en) * 2020-06-18 2020-10-09 上海高德威智能交通系统有限公司 Target detection method, device and equipment
CN111768246A (en) * 2020-06-30 2020-10-13 创新奇智(北京)科技有限公司 Data processing method, model establishing device and electronic equipment
CN112183191A (en) * 2020-08-18 2021-01-05 天津大学 Method for detecting bird nest in power transmission line based on countermeasure training
CN112395959A (en) * 2020-10-30 2021-02-23 天合云能源互联网技术(杭州)有限公司 Power transformer fault prediction and diagnosis method and system based on audio features

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
AL-MASNI M A等: "Simultaneous detection and classification of breast masses in digital mammograms via a deep learning YOLO-based CAD system", 《COMPUTER METHODS AND PROGRAMS IN BIOMEDICINE》 *
H. XU等: "Performance Comparison of Small Object Detection Algorithms of UAV based Aerial Images", 《2020 19TH INTERNATIONAL SYMPOSIUM ON DISTRIBUTED COMPUTING AND APPLICATIONS FOR BUSINESS ENGINEERING AND SCIENCE (DCABES)》 *
I. CHYRKA等: "1D direction estimation with a YOLO network", 《2019 EUROPEAN MICROWAVE CONFERENCE IN CENTRAL EUROPE (EUMCE)》 *
阮激扬: "基于YOLO的目标检测算法设计与实现", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780063A (en) * 2021-07-27 2021-12-10 深圳泰豪信息技术有限公司 Photovoltaic operation and maintenance control method based on video intelligent analysis
CN113486866A (en) * 2021-09-06 2021-10-08 南京天朗防务科技有限公司 Visual analysis method and system for airport bird identification

Also Published As

Publication number Publication date
CN113076860B (en) 2022-02-25

Similar Documents

Publication Publication Date Title
CN108334848B (en) Tiny face recognition method based on generation countermeasure network
CN110348319B (en) Face anti-counterfeiting method based on face depth information and edge image fusion
CN109670429B (en) Method and system for detecting multiple targets of human faces of surveillance videos based on instance segmentation
CN108898137B (en) Natural image character recognition method and system based on deep neural network
CN107944427B (en) Dynamic face recognition method and computer readable storage medium
CN111445459B (en) Image defect detection method and system based on depth twin network
CN111241975B (en) Face recognition detection method and system based on mobile terminal edge calculation
CN109767422A (en) Pipe detection recognition methods, storage medium and robot based on deep learning
CN108647625A (en) A kind of expression recognition method and device
CN113076860B (en) Bird detection system under field scene
CN113011357A (en) Depth fake face video positioning method based on space-time fusion
CN112287896A (en) Unmanned aerial vehicle aerial image target detection method and system based on deep learning
CN110599463A (en) Tongue image detection and positioning algorithm based on lightweight cascade neural network
CN111612759B (en) Printed matter defect identification method based on deep convolution generation type countermeasure network
CN112784712A (en) Missing child early warning implementation method and device based on real-time monitoring
CN112861785A (en) Shielded pedestrian re-identification method based on example segmentation and image restoration
CN113591825A (en) Target search reconstruction method and device based on super-resolution network and storage medium
Vaishak et al. Currency and fake currency detection using machine learning and image processing–an application for blind people using android studio
CN112488165A (en) Infrared pedestrian identification method and system based on deep learning model
CN115797970B (en) Dense pedestrian target detection method and system based on YOLOv5 model
CN111199050B (en) System for automatically desensitizing medical records and application
CN116521917A (en) Picture screening method and device
CN116758421A (en) Remote sensing image directed target detection method based on weak supervised learning
CN116206373A (en) Living body detection method, electronic device and storage medium
Selim et al. Image Quality-aware Deep Networks Ensemble for Efficient Gender Recognition in the Wild.

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CB03 Change of inventor or designer information

Inventor after: Lu Yu

Inventor after: Ramenco

Inventor after: Xiao Weikang

Inventor after: Zhong Zhifang

Inventor after: Yang Shixiong

Inventor after: Wang Yan

Inventor before: Ramenco

Inventor before: Xiao Weikang

Inventor before: Zhong Zhifang

Inventor before: Yang Shixiong

Inventor before: Wang Yan

CB03 Change of inventor or designer information