Unmanned aerial vehicle route safety detection method based on improved Faster RCNN

Abstract

The invention provides an unmanned aerial vehicle route safety detection method based on an improved Faster RCNN, and belongs to the technical field of image processing. According to the method, firstly, VGG16 is used as a convolutional neural main network to establish a Faster RCNN basic model, image data of an unmanned aerial vehicle route are input into the model, an image feature map is obtained through convolution calculation, and different candidate areas are mapped to obtain feature maps of the candidate areas; and then introducing a minimum entropy potential model to process the feature map of the candidate region. And finally, performing target identification and detection on the low-altitude unmanned aerial vehicle in the unmanned aerial vehicle air route. According to the unmanned aerial vehicle air route safety detection method based on the improved Faster RCNN, the local minimum problem can be effectively avoided, the calculation burden of a candidate area is reduced, good target detection performance is achieved, and a new technical means is provided for unmanned aerial vehicle air route safety detection.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for detecting the safety of an unmanned aerial vehicle route based on an improved Faster RCNN. Background Art

[0002] In recent years, small drones such as "low, small, and slow" have developed rapidly and are widely used in information communication, environmental monitoring, express logistics and other fields. While bringing convenience to our lives, they also pose a great threat to public safety. At present, the commonly used detection technologies for drone route safety include radar detection, acoustic detection, and radio frequency detection, but these methods often require expensive detection equipment, and there are certain shortcomings in the detection effect in complex environments. Therefore, low-altitude drone route safety detection technology based on machine vision has a good development prospect.

[0003] In the process of acquiring and perceiving image information using computers, target detection is an important and basic research direction. Many studies rely on it. The important criteria for evaluating target detection in images include target recognition and target positioning, that is, in addition to accurately identifying low-altitude drone targets in images, their positions must also be located. The current mainstream research method in the field of target recognition and detection is convolutional neural network (CNN).

[0004] CNN has a good effect on the recognition of single background targets, but it is difficult to cope with more complex images. In this case, it is necessary to add a target detection algorithm to detect the required targets first and then carry out the recognition task. Scholars have proposed a variety of research methods for target detection based on machine vision. Currently, the commonly used target detection algorithms include anchor-free detection algorithms, two-stage detection algorithms including R-CNN, Faster RCNN, and single-stage detection algorithms including YOLO. However, the above methods still have some shortcomings in the field of low-altitude drone route safety.

[0005] Two-stage target detection algorithm: first generate target candidate frames, and then use convolutional neural networks to identify targets in the candidate frames. The key lies in the selection of candidate regions. Often, there are many generated target anchor frames, but few target individuals. The parameters introduced by the anchor frames make the target detection process more complicated.

[0006] Single-stage object detection algorithm: This method mainly regresses bounding boxes and categories, which can better meet real-time requirements. However, although its algorithm runs faster, it still faces problems such as loss of accuracy. Researchers have also continuously improved single-stage algorithms represented by YOLO.

[0007] In summary, many current technical means have considered target detection methods based on machine vision, but they still have shortcomings in terms of real-time performance and detection accuracy required for the safety of low-altitude UAV routes. Summary of the invention

[0008] The purpose of the present invention is to provide a UAV route safety detection method based on improved Faster RCNN, which can effectively avoid the local minimum problem, reduce the calculation burden of the candidate area, have better target detection performance, and provide a new technical means for UAV route safety detection.

[0009] To achieve the above object, the present invention provides a UAV route safety detection method based on improved Faster RCNN, comprising the following steps: Step S1, using VGG16 as the convolutional neural network, establishing the Faster RCNN basic model, inputting the image data of the drone route into the Faster RCNN basic model, obtaining the image feature map through the convolution calculation of the neural network, and mapping different candidate areas to obtain the feature map of the candidate area; Step S2: introduce the minimum entropy potential model to process the feature map of the candidate area: Divide the feature map of the candidate region into multiple different target clusters; Evaluate the global entropy value of the divided target clusters, find the cluster system with the smallest global entropy value and activate it; Calculate the local entropy value in the activated target group, select the candidate box with the smallest local entropy value, and generate the feature map of the candidate box after being processed by the minimum entropy potential model; Step S3: perform target recognition and detection on low-altitude drones in the drone route.

[0010] Preferably, in step S3, target identification and detection of low-altitude drones in the drone route includes: Use non-maximum suppression to handle redundant candidate boxes: ; in, represents the candidate box obtained after processing, represents the softmax function; Represents the softmax function used to generate the candidate box. represents the output of the fully connected layer, Represents the output result of the fully connected layer of the target detection box, Represents vector point-by-point product operation; The feature map of the candidate area processed by the minimum entropy potential model is processed, and the category probability value of the candidate area is solved through the fully connected layer and the softmax function: ; in, Indicates target category, The softmax function under the feature map, Indicates target category, Under the feature map The output of the fully connected layer of each category is represents exponential operation, Indicates target category, Under the feature map Category fully connected layer output; The feature map of the candidate area processed by the minimum entropy potential model is processed, and the probability value of the candidate area corresponding to each recognition target is solved through the fully connected layer and the softmax function: ; in, Indicates Candidate boxes, The softmax function under the feature map, represents the candidate region after processing by the minimum entropy potential model, Indicates Candidate boxes, The fully connected layer output of the target detection box of the feature map, Indicates Candidate boxes, The fully connected layer outputs the target detection box of the feature map.

[0011] Therefore, the present invention adopts the above-mentioned UAV route safety detection method based on improved Faster RCNN, and the beneficial technical effects are as follows: (1) Avoiding the local minimum problem: The minimum entropy potential model is used to overcome the difficulty that the Faster RCNN algorithm usually faces in solving non-convex problems during the detection process. Based on the clique system, the minimum entropy potential model reduces the local randomness in the target detection process by means of the target clique faction division, global minimum entropy optimization, and local minimum entropy optimization process.

[0012] (2) Advantages in target detection performance: Methods such as obtaining target candidate regions based on the minimum entropy potential model and using non-maximum suppression to screen candidate regions reduce the computational burden of candidate regions. Therefore, the Faster RCNN algorithm improved based on the minimum entropy model has better target detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a UAV route safety detection method based on improved Faster RCNN; Figure 2 This is the minimum entropy potential model flow chart. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0015] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0016] Embodiment 1 1. Establish the Faster RCNN basic model.

[0017] In this embodiment, the convolution neural network of Faster RCNN uses VGG16 to increase the network depth and effectively improve the performance of deep learning. The convolution layer and the fully connected layer are related to the parameter settings of the entire model, which need to be adjusted continuously during the training process. The bounding box is given to the image of the low-altitude small drone data set, and then it is generated into a candidate box through the selective search strategy.

[0018] 2. Introduce the minimum entropy potential model.

[0019] The minimum entropy potential model is a learning model based on the group system and the concept of entropy. The model mainly includes the target group faction division, global minimum entropy optimization and local minimum entropy optimization process.

[0020] 201. Division of factions within the group.

[0021] The candidate regions of the image are extracted through the minimum entropy potential model, and the candidate regions are divided into factions and divided into multiple different target groups.

[0022] In the learning phase, target candidate boxes are first generated for each image, and then the features of these candidate boxes are extracted in the ROI pooling layer above the convolution layer. The optimization process is continuously carried out using the learning algorithm. The forward process continuously learns and selects target instances, while the reverse process needs to continuously optimize the parameters of the network structure to reduce the loss value during the training process.

[0023] ; in, is the total number of target categories, Indicates image-level class prediction scores, Indicates image-level class prediction probabilities.

[0024] like Figure 2 As shown in the figure, the image is processed by the convolutional neural network, and then the processes of clique division, target group discovery and target location are repeated. In the process of deep network learning, the desired objective function (target score) needs to be continuously optimized during the clique division process to solve the non-convex problem faced in the network learning. Finally, the recognition and detection results are output through the image classifier and target detector.

[0025] 202. Target cluster discovery based on global minimum entropy.

[0026] Treat all target clusters as a whole, evaluate the entropy value within the whole range, find the cluster system that makes the global entropy value the smallest, and activate these clusters, ignoring other clusters with larger entropy values. Therefore, the cluster with the lowest global entropy value can be found and used as the desired target cluster.

[0027] 203. Target positioning based on local minimum entropy.

[0028] In step 202, the local entropy value of the target cluster is calculated in the cluster system activated based on the global minimum entropy. The size of the entropy value is also used to suppress the candidate areas with too much redundancy, and then the candidate area with a smaller entropy value is selected based on the local minimum entropy.

[0029] 3. Target recognition and detection process.

[0030] According to the two branches of image classification and target detection, the softmax function is processed along the image target recognition and target positioning of the low-altitude small drone dataset, thereby realizing the target classification and detection frame regression process in the image. Although the two branches focus on different directions, they both have the functions of recognition and positioning, but the softmax layer processes in different directions.

[0031] 301. Non-maximum suppression.

[0032] The candidate region with a high score is most likely the region you want to get in the end. In order to minimize redundancy, NMS (non-maximum suppression) is added to minimize redundant candidate boxes and obtain the region of interest.

[0033] ; in, represents the candidate box obtained after processing, represents the softmax function; Represents the softmax function used to generate the candidate box. represents the output of the fully connected layer, Represents the output result of the fully connected layer of the target detection box, Represents vector point-by-point product operation; 302. UAV target recognition branch.

[0034] The drone target recognition branch is to solve the category probability value of the candidate area. Based on the image labels in the dataset, the region of interest obtained in the above steps is pooled, and the feature map of the candidate area is processed by a fully connected layer and output The softmax function is used to normalize the categories (target recognition process), and the result is equivalent to the probability arrangement of target recognition in a certain area.

[0035] ; in, Indicates target category, The softmax function under the feature map, Indicates target category, Under the feature map The output of the fully connected layer is represents exponential operation, Indicates target category, Under the feature map Category fully connected layer output; 303. UAV regional positioning branch.

[0036] The drone area positioning branch is to solve the probability value of the candidate area corresponding to each identified target. The feature map of the obtained candidate area is processed by the fully connected layer and the output is The softmax function is used to perform normalization along the candidate regions (detection box regression process) to obtain the positioning probability arrangement of a certain identified target in all candidate regions.

[0037] ; in, Indicates Candidate boxes, The softmax function under the feature map, represents the candidate region after processing by the minimum entropy potential model, Indicates Candidate boxes, The fully connected layer output of the target detection box of the feature map, Indicates Candidate boxes, The fully connected layer outputs the target detection box of the feature map.

[0038] It is worth noting that the contents not elaborated in detail in the present invention are all prior art and are well known to those skilled in the art.

[0039] Therefore, the present invention adopts the above-mentioned UAV route safety detection method based on the improved Faster RCNN, which can effectively avoid the local minimum problem, reduce the calculation burden of the candidate area, and has better target detection performance, thus providing a new technical means for UAV route safety detection.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A UAV route safety detection method based on improved Faster RCNN, characterized in that: The following steps are involved: Step S1, using VGG16 as the convolutional neural network, establishing the Faster RCNN basic model, inputting the image data of the drone route into the Faster RCNN basic model, obtaining the image feature map through the convolution calculation of the neural network, and mapping different candidate areas to obtain the feature map of the candidate area; Step S2: introduce the minimum entropy potential model to process the feature map of the candidate area: Divide the feature map of the candidate region into multiple different target clusters; Evaluate the global entropy value of the divided target clusters, find the cluster system with the smallest global entropy value and activate it; Calculate the local entropy value in the activated target group, select the candidate box with the smallest local entropy value, and generate the feature map of the candidate box after being processed by the minimum entropy potential model; Step S3: perform target recognition and detection on low-altitude drones in the drone route.

2. According to claim 1, a method for detecting the safety of a UAV route based on an improved Faster RCNN is characterized in that: In step S3, target recognition and detection of low-altitude drones in the drone route are performed, including: Use non-maximum suppression to handle redundant candidate boxes: ； in, represents the candidate box obtained after processing, represents the softmax function; Represents the softmax function used to generate the candidate box. represents the output of the fully connected layer, Represents the output result of the fully connected layer of the target detection box, Represents vector point-by-point product operation; The feature map of the candidate area processed by the minimum entropy potential model is processed, and the category probability value of the candidate area is solved through the fully connected layer and the softmax function: ； in, Indicates target category, The softmax function under the feature map, Indicates target category, Under the feature map The output of the fully connected layer of each category is represents exponential operation, Indicates target category, Under the feature map Category fully connected layer output; The feature map of the candidate area processed by the minimum entropy potential model is processed, and the probability value of the candidate area corresponding to each recognition target is solved through the fully connected layer and the softmax function: ； in, Indicates Candidate boxes, The softmax function under the feature map, represents the candidate region after processing by the minimum entropy potential model, Indicates Candidate boxes, The fully connected layer output of the target detection box of the feature map, Indicates Candidate boxes, The fully connected layer outputs the target detection box of the feature map.

Images