Analysis of Drone Target Search Performance According to Environment Change

Abstract

In recent years, interest in drones has grown, and many countries are developing them into a strategic industry of the future. Drones are not only used in industries such as logistics and agriculture but also in various public sectors such as life rescue, disaster investigation, traffic control, and firefighting. One of the most important tasks of a drone is to accurately identify targets in these applications. Target recognition may vary depending on the search environment of the drone. Therefore, this study tests and analyzes the drone's target recognition performance according to changes in the search environment such as the search altitude and the search angle. In addition, we propose a new algorithm that improves upon the disadvantages of the Haar cascade method, which is the existing algorithm that recognizes the target by analyzing a captured image.

1. INTRODUCTION

Drones are a type of unmanned aerial vehicle capable  of flying and being remotely controlled by radio. The drone market has grown rapidly as drones are used for a variety of purposes such as disaster safety, broadcasting, transportation, and leisure sports. This is particularly true in the public sector where they serve a variety of uses for the public interest, such as military reconnaissance, disaster information and relief, lifesaving, trafficconcerns, crop growth, environmental protection, geographic information acquisition, and logistics [1]. In most of these drone applications, accurately recognizing the target is one of the most importanttasks of the drone. Especially in disaster scenes and accident situations, it is important to discern the target precisely because it is directly connectewith human life [2-3]. The probability of a drone 'ssuccessful target search depends on the its searchenvironment, There are various factors that may impact the drone's search environment. When adrone searches for a target on the ground, thesearch altitude, the search angle, and the camerasensor performance are major factors that influence its & nbsp; perception of the surrounding environment [4]. Of these, the search altitude and the search angle   have the most affect the search performance. Because the search altitude and the search angleare changing during flight, the success probability of the target search can vary greatly.

Therefore, we analyze a drone's target search success according to various search altitudes and search angles through experiments and study the optimal drone flight environment. In this study, we used the human recognition module of the OpenCV library as a method to analyze the existence of the target by processing the images acquired by the drone. In particular, we explore ways to improvethe use of the high-altitude aerial reconnaissance cascades (known as Haar cascades) module whichis known for its superior performance, to increasethe likelihood of a target search success.

This paper is organized as follows: In Section II we investigate related research and analyze therelevant literature to extract the features of this study. In Section III we describe the environment of the experiment conducted in this study and examine   the theoretical background of the experiment. In Section IV we explain the experiment conducted to examine recognition performance of the droneaccording to the environmental changes and analyze  the results. Finally, in Section V, we discussthe contributions and conclusions of this study

2. RELATED WORKS

Target search methods for drones include image processing, communication signal processing, and probability-based target search [5]. Liu [6], Wang[7], and Mejias [8] are representative studies of image   processing-based object search methods. In these methods, the drones detect the target through the camera sensor and store the detected imagedata in an on-board storage device. The acquiredimage is transmitted to the base station through wireless networks and the transmitted image is analyzed using an image processing algorithm. The existence of the target is then examined and discriminated.

Arora [9], Costa [10], and Jawhar [11] examinerepresentative signal processing-based target search methods. In these methods, the dronessearch for a signal to detect the presence of the target. However, in most cases, the targets aregenerally not producing signals. An example of such a case is a person hurt at a disaster site &nd ash; it is likely that there is no electronic signal beingemitted by the victim.

The technique of probing the target by probabilistic   methods is a useful approach in the autonomous & nbsp; flight of the drone. In this method, when adrone cannot transmit a sensed image data to a  remote site, the drone automatically discovers the target. In this method, the drone narrows the areawhere the target is likely to exist based on the initial  information about the search area, thereby increasing   the possibility of finding the target. Studies on this method include Ha [5], Chung [12], Waharte [13], and Symington [14].

There are few studies dealing with environmental factors & nbsp; to increase the likelihood of drones finding the targets including, that by Ha [5,15] and Syming ton [14]. The subject of [5] is collaboration of drones to increase the likelihood of success of the drone’s target search, and [15] investigates thesearch altitude control of drones to improve thesearch performance.

On the other hand, in finding the target by the drones, it is important to process the image information acquired from the search area. The algorithm for recognizing the object in   the captured image is similar to the conventional image processing technique. If the target is a person, an image   processing library such as OpenCV (Open Source Computer Vision) can be used. OpenCV is an image   processing library focused on real-time image processing [16]. In addition, the image processing algorithms can extract or process information in a target image to be recognized by using an existing   picture or a photograph. OpenCV is used in many computer vision applications, including facedetection and recognition and object recognition. In this study, the OpenCV library is used in an algorithm to   process images acquired from a searcharea and recognize objects. Specifically, we use OpenCV's cv2 library for image processing. Thealgorithms used in this study, Haar cascade [17]and histograms of oriented gradient (HOG) cascade  [18], use learned data to increase object rec-ognition probability. The Haar cascade module provides high-speed detection but with a relatively low recognition success rate. On the other hand, HOG cascade is advantageous in that, while the object detection speed is low, the recognition success & nbsp; rate is relatively high. In this study, we usethe Haar cascade method to detect the target (person). In particular, we improve the Haar cascade   method of applying the algorithm to the entirebody of a person, and reduce the false recognition rate of the target by applying the image processing algorithm separately to the upper body and the lower body.

Therefore, the contributions and features of ourstudy are as follows. First, the experiment confirms & nbsp; that the target discovery probability varies according to changes of the drone's target searchenvironment, given by the search altitude and search angle. Second, we use an improved Haarcascade method for recognizing the target by   processing the image acquired through the camera  sensor by the drones.

3. TARGET SEARCH PERFORMANCE
ACCORDING TO ENVIRONMENTAL
CHANGE

The target search performance of a drone canvary depending on changes in environmentafactors. Of these factors, this study focuses on thesearch altitude and the search angle and observesthe associated changes in search performance. In this section, we describe our experiment and determine   the drone’s target search performance as a function of the environmental changes, and then we explain the image processing algorithm used to determine the existence of the target.

3.1 Experiment to Determine Target Search
Performance

In this experiment, we observed changes in the success rate of the target search by changing thesearch altitude and the search angle of the drone. Table 1 shows the experimental environment for the target search performance test. The experimental  drone used is the DJI Phantom 3 model, and the experiment was conducted on the athletic field as shown in Fig. 1. The targets in this experiment   are three people located on the playground.

We experimented by changing the distance and angle from the target to the drone. Experiments were performed with the straight-line distance from the target to the drone (c) set at 10 and 20 m, and experiments are performed at an angle (θ ) of 30, 45, 60, 75 and 90° for each distance. Let ' A' be the target and 'B' be the drone as shown in Fig. 1; 'c' is the straight-line distance from the target to the drone and ∠A is the angle (θ) to betested. Therefore, the search height (a) of the dronecan be obtained as shown from Eq. (1). In practice, since the drone used in the experiment provides thevertical distance ‘a’ and ground distance ‘b’, we usethis data to find the distance ‘c’ between the drone and the measurer. The search angle can be obtained & nbsp; according to Eq. (1).

Search altitude (a) can be found using the Pythagorean identity:   ×sin (1)

Table 1. Environment of the target search experiment

Fig. 1. A drone performance test scene.

3.2 Image Processing Algorithm for Target
Recognition

In Analyzing the data our drone collected requires & nbsp;image processing. For this study, our targets were   human and there are extensive products for human target recognition. We used the existing   algorithms to available from the cv2 library provided by OpenCV. The cv2 library can be used to easily apply data that has been learned in advance   and it provides classification algorithms for human recognition, namely, the Haar cascade and HOG cascade classifiers. The Harr cascade is the most traditional and representative object recognition   algorithm. This algorithm recognizes objects by sequentially applying several detectors, and is relatively fast because image processing for the final   detection is performed only on images filtered by a detector [19]. On the other hand, the HOG cascade   technique uses the cascade technique to generate the  histogram [20]. It sets various blocks forobject recognition, calculates the HOG, and boosts meaning ful blocks. This method has a relatively high recognition rate while the detection speed isslow [21,22].

The speed of the drone’s search for targets isimportant; i.e., it is important to analyze the imagescaptured by the drone quickly and judge the existence  of the target. Therefore, in this study, we choose the Haar cascade algorithm which has the advantage of rapidity analyzing the image. The Haar cascade algorithm included in the existing OpenCV library is an algorithm that recognizes the entire human body. When this algorithm is applied to a drone’s target search experiment, the false detection   rate for perceiving people as human being sis high. Therefore, we intend to reduce this high false detection rate as follows. When the Haar cascade   algorithm is applied to the experiment image, if it is recognized as a human object, it is divided into an upper body and a lower body, and anotherclassifier is applied to each of them to recognizea person. If either is recognized as a person, the image is recognized as a human object. Fig. 2  compares the three algorithms used in the experiments in this study. As shown in the figure, the ‘ Haar_body ’ method is a method of recognizing an objecusing the whole body as a unit object, and the & lsquo; Haar_face ’ method is a method of recognizing anobject based on a face of a body. The proposed method is to recognize the object by separating the upper and lower body parts of the body. The details of the proposed algorithm are shown in Fig. 3

Fig. 2. Improved target detection algorithm.

Fig. 3. Improved target detection algorithm.

4. EXPERIMENTS AND RESULTS FOR
TARGET SEARCH PERFORMANCE

In the first experiment, the distance between the drone and the target is 10 m. At that time, thesearch altitude is 10 x sinθ according to Equation (1). The search angle is then changed from 30° to 90 & deg; . Images taken at each search angle are stored and analyzed through the OpenCV library. (a) of Fig. 4 shows the image processing results when the straight-line distance between the drone and the target is 10 m and the search angle is changed from 30° to 90°. The left side of the figure shows the result of processing with the existing algorithmand the right side shows the result of processingusing the improved algorithm.

The second experiment is carried out with aseparation of 20 m between the drone and the target. At that time, the search altitude becomes 20 x sinθ by Equation (1). As in the first experiment,   the test is performed while changing thesearch angle from 30° to 90°. (b) of Fig. 4 shows the image processing results when the distance   between the drone and the target is 20 m and thesearch angle is changed from 30° to 90°. The left and right columns show the results of analyzing the images taken at each search angle using the existing algorithm and the improved algorithm, respectively. Objects recognized by the existing algorithm are shown with a blue border, and a full body object recognized by the improved algorithmis shown with a blue rectangle border. The upperbody recognized as an object is represented by agreen color, and the lower body is represented by a red square.

Fig. 5 shows the results of the first experiment. As shown in the figure, when the existing algorithm is used, the target recognition rate is high, but the false recognition rate (other objects arerecognized as people) is also high. To solve theseproblems, we proposed the algorithm shown in Fig. 3. As shown in the figure, when the proposed improved & nbsp; algorithm is applied, the recognition rate for the object is somewhat low, also the false recognition   rate can be greatly reduced. In Fig. 5, we can see that the false recognition rate decreases as the search angle increases. This is because the 

images obtained at each search angle are different from each other. However, it is proven that the  existing algorithm has a high false recognition rate. Fig. 6 shows the result of analyzing the secondexperiment. The results are similar to the firstexperiment. When using the existing algorithm, therecognition rate is relatively high, but the false  recognition rate is high. On the other hand, when the improved algorithm is applied, the false recognition rate is greatly reduced. We conclude that the  recognition rate decreases when the search distance increases or the search altitude increases. In both experiments, we confirmed the following facts. First, it is possible to significantly reduce the false recognition rate when the improved algorithmis used than when the existing algorithm is used. Second, as the search angle increases, the recognition   rate decreases somewhat. Third, as thesearch altitude increases, the recognition rate decreases somewhat.

Fig. 4. Detected The image processing results when the search altitudes are 10 x sinθ and 20×sinθ.

Fig. 5. Experimental results when the search altitude is sinθ× 10 m.

Fig. 6. Experimental results when the search altitude is sinθ x 20 m.

5. CONCLUSION

In this study, we investigated the changes of target search performance caused by changes of the search environmental factors, such as the drone 's search distance and search angle. We also proposed a new algorithm that improves the falserecognition rate of the Haar cascade algorithm provided by OpenCV. Experimental results show that the drone search performance changes according to the search angle and the search altitude. Especially, it is confirmed that the object recognition rate decreases as thesearch altitude increases and the search angle increases. In addition, we confirmed that our algorithm improved the Haar cascade and is more effective than the  existing algorithm in reducingfalse object recognition