Scribd
Upload
42 views

Caaaaa

sad

Uploaded by

Frkan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
42 views

Caaaaa

sad

Uploaded by

Frkan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 8
‘ah oral of gern = 20215) 193.20 ms Image processing based autonomous landing zone detection for a multi-rotor drone in ‘emergency situations Veysel Turan ©, Ercan Avgar"\®, Davood Asadihendoustani®®, Emine Avsar Aydin “akrova Universi: Focly of Engineering, Deparementof lectrical and Electronics Engineering, Adana. Turkey “Adana Arson Trg Ste ce Techeogy Unies Foal of roma nd rons, Department of Arena Erginering Adon, Turkey ‘Anton landing Rpt ste and relabilty improvemests an mprtant research ue imei applitions Image processing ‘Multerotor drones are wnersble to motor fares leeding to potential safe operations Objet detecton or clisons. Therefore researchers are working on autopomeus landing mses to safely av ‘ecover an andthe ay drone non a desredlandingareaInsuch aca. aaulable nding, ane shouldbe detected rapidly infor emergency landing Majority ofthe works related wth ‘Sttonomous landing utlze a marker and GPS signals to detect nding se Ins work we ‘ropoee landing system framework that involves ony the procesing of mages taken fom ‘he onboard camer ofthe vec Fis the objects the image are determined itring Sh edge detection algorithm, then the not sufable andng zone searched. The ares hat |stree from obstaces and closet othe center ofthe images defined as the most immediate fod slable landing zone. The method has been tested on 25 mages taken rom diferent ‘sits and performance hae been evalsatd in term runtime on single boa compet {hd detection precision and recall values The average measured runtine 24923 seconds Sd 100% of precision and recall values are achieved for the images taken from 1m and 2m. ‘The males pression andrea valine are 79.1% 30 4. INTRODUCTION ‘The use of Unmanned Aerial Vehicles (UAV) has increased. at an unpredictable rate in recent years Although these devices have been used particulary in nilitary applications fora long time, their use in non- nilitary applications sich as fre extinguishing (Aydin et al 2019), meteorological research (Martin ef al 2010; PropotoAV 2019), exploration (Jakob et al, 2016 Heincke et al. 2019) and. agricultural activities (Weroustracte 2015) have become very” widespread nowadays. Quadrotors are the most common devices to Use among UAV types due to their uncomplicated ‘mechanical structures. Quaérotors can fll 3D mation tracking requirements with various technical systems such as Global Positioning System (GPS), ultrasonic etection, angular velocity” sensors and linear accelerometers (Zhao and Wang 2012). Despite the use ofthese integrated systems and sensors, the control of ‘the quadrotors is stil one of the most dificult isues, but 2% respective research and development activities continue in many esearch centers (Hoffnann et al. 2007; Zhao and Wang 2012). The vast malonty of esearch activities address control issues during lying. The ultimate goal of UAV systems isto reach fully autonomous operations (Kim nd Sukkarieh 2002), ‘One of the most important issue for UAV is autonomous landing in motion. Meanwhile, demands for ‘tomatic landing of drones ona defined position, safely, accurately, and without human's intervention, inerease everyday. tis aways posible fora UAV to unavoldably face emergencies during fight, such as engine fallure, Interruption of data link from the ground and other unexpected accidents (strong wind, rain, ete). Thus forced landing measures should urgenty be adopted in such situations Methods of forced landing such as parachute and other ight termination systems can cause damage onthe body of the muli-rotor (Fitzgerald et 2008) n addition, GPS signals are highly susceptible to be interrupted especially at lower altitudes (Lee eta (cctidoacoueaun) opines one—Tvs- ons feeserer oe oa nea ‘Raussen se oe a ‘iar tengning S) 850 Ree 29/05/3120 Acre 2007/2820 ‘arc oral ofEngerng- 202 5): 188200, 2012; Ho 2017), and in indoor environments Furthermore, GPS signals are controled by other nations, whlch causes vulnerability issues. For Instance, In 29th (October 2018, the GS jamming that aused 46 drones to plummet during a display over Victoria Harbour caused Atleast HKSI milion (USS127,500) of damage, according toa senor official from the Hong Kong Tourism Board (Liteye 2018). As a result, some alternative methods ‘were adopted to minimize the damage on that UAVs by enabling them to autonomously finda safe area suitable for landing. Some studies about forced landing in ‘emergency situations for UAVs were presented in indoor environment without using GPS (Nemati etal. 2015) In the past years, in design of emergency landing systems, effective algorithms of machine earning suchas Support Vector Machines (SUM) and Artificial Neural [Networks (ANN) were utilized in combination with dighal image processing techniques for selection of appropriate landing site (Guo et al 2014, Lunghl et al 2015} Its observed thatthe above mentioned machine learning algorithms have different performance constraints, eg. SVM is complex and requires huge computational power, whereas, ANN requires large training dataset which corresponds to greater training time, Due to these constraints, both algorithms cannot inet the rapidly changing requirements for landing area selection in emergency fight conditions. Fer instance, in 4 previous study about detection of forced landing sites, 901 images were used to train and test ofan ANN mode (Fitzgerald and Walker 2005). Because the images fed to the ANN model were not representative of the training set the fnaleassifcation accuracy is very low when ‘compared to training accuracy (La etal. 2013). Thus that ‘Station prevents from being completely reliable. Image processing isan appropriate and reliable way to find safe landing sites n case of an emergency. This is senerally accomplished by detection of objects in the mages taken from UAVs. Most ofthe related studies are about detection of some “marker” in an image (Barts et al 2014; Cabrera-Ponce and Martinez-Carranza, 2017; Sani and Karimian 2017) which represents the desired landing spot Utilization ofa marker may not be a feasible way for emergency landing conditions. Since markers ‘an be fr from the UAV then wil talke time to find the smarkers. For instance, the recent demonstration of package delivery using an UAV by Amazon shows the feasibility ofan UAV sending a package to Is consumer (AMAZON 2017). A marker which Is placed by the consumer on the ground is used to allow the UAV to land safely. In some circumstances where the marker is ‘unavailable or inappropriately placed, these vehicles feed fo be able to sense and avold the surrounding ‘objects in the environment and. perform a. smooth descent automatically. Without this capability, the safety of the surrounding” animals, humans, and. property cannot be ensured Additionally, in some emergency situations, the UAV needs to land a8 quickly as possible, ‘Therefore, i is essential to select a safe landing spot automatically and emergenty, without depending on external ystems. Motivated by the above reasons, in this paper, an Image processing method for object detection is proposed The method is developed to work on the images taken froma drone's onboard camera The aim is to rapidly detect’ a suitable landing. zone in an unstructured and unknown environment. The method Initially proposes a candidate landing zone a the center ofthe image. Ifthe initial proposal isnot suitable then new candidates inthe neighboring area are evaluated until an available spot is detected (Fig. 1). The suitability ofa spots determined by existence af an objctinsie it ‘The object detection is accomplished by means of several image processing methods including edge detection, color processing, morphologic operations and thresholding The major advantages ofthe method are (0) no requirement for a marker and (i) no need for huge mount of data for training.a model, 2, METHOD. 2.1. The Landing System Framework. ‘The use of multtotor drones has undeniably \nereased in the past decade, both in the military and ivan applications, this raising a number of vital unsolved issues including safety and reliability. Engine ‘malfunction or failures are among the common faults in ‘ultrotor drones, which apparently endanger the Arone and the people's safety onthe ground In order to Inerease fight safety and reliability of drones, researchers are working on automation enhancement 19 safely recover the impaired drone (Loper-Franco eta 2017, Mazeh etal 2018; Nguyen etal 2019). ‘There are several challenges related. with safe recovery or landing of impaired arial vehicles. Majority ofthese challenges are about obstacle detection, suitable landing site detection/seection, fault detection and Identification, characterizing the alreraft’s new inematie constraints, trajectory planning and control of the faulty aircraft on the landing trajectory. To cover these challenges, an emergency landing system has been proposed acording to Fig 1. fault oF fallure scenarios where continuation of ight is not possible or endangers the ight safety the emergency fight systems trigaered to recover drone's stability and safely land the drone on ‘stable landing ste ‘The emergency landing system Is translated to an architecture consisting of various subsystems that are Capable landing faulty droneto a desired landing site along a designed trajectory without colliding to any human or animal. The architecture autoniomosiy detects objects 2s. well as possible landing. sites, determines the most suitable landing st, develops the landing trajectory based on new kinematic and dymane constraints of impaired drone, and controls it to the Tanding site, using onboard camera data and other fommon sensor Information lke IMU (Inertial Measurement Uni) 22. Landing Zone Detection Emergency landing mode of the UAV is activated whenever a fault in any of the motors is detected, Then the onboard camera is immediately triggered to take 2 topzview photo ofthe ground in perpendicular direction. ‘This mage isthe eld of view (FOV) of the UAV and it constitutes the search space for finding suitable landing 196 ‘Tue ura of Bgierng- 2021 519-200 zone. In such an emergency situation, locating the Appropriate 2one in a short tie is very important ‘Therefore the algorithm starts checkingthe stability of center of the mage which denotes the closest area to the TUAV. Ifthe image center i occupied by an objec than t is labeled as “negative” and neighborhood ofthe image The distance of the center checked for” suitability neighborhood tothe image center is gradually increased nti vacant spot found and that spots labeled tobe positive”. Nex, the coordinates ofthis “positive” spots sent to the fight controller to initialize the autonomous landing process. The steps for emergency landing are presented in Fig, 2 Camera Lev VSAAFP Sub-Systems Failure Figure 1, Emergency landing system architecture and subsystem Talia 1-0 dmnce trom the center Propose a landing erlap with ay objet Jeane as "negative of distance proposed? Figure 2. Suable Landing site detection flowchart 2.2.1. Object detection in FOV images ‘The object detection starts with converting the FOV Image into a grayscale image. This is necessary for preparing the image for canny operator, which ts 2 ‘method for edge detection to find objects on image. The Canny edge detector is an edge detection operator that tases a miltestage algorithm fo detect a wide range of felges in images (Canny 1986). The advantage of sing Canny edge detection technique over other well-known fvlge detection algorithms fs that i gives better results ven in nosy conditions (Kumar etal. 2015), 195 ‘The edge detection operator has four steps. ‘+ Smooth an image with Gaussian filter + Calculate gradient magnitude and_ gradient direction, ‘+ "Non = maximum suppression” to ensure the desired edge with one singe pixel with, ‘+ Determine two threshold valtes, and then select possible edge points and trace edges ‘The frst step in canny edge detector algorithm Isto delete the noise inthe frames by applying a Gaussian fiter. In Canny algorithm, the Gaussian finction is applied to smooth the image prior to eige detection. The filtering or smoothing operation actually supports two purposes. The frst one is noise eect reduction prior to the detection of pixel intensity changes. The second purpose is setting the resolution of scale at which Intensity changes are to be detected (Chen et al. 2014) These two purposes are necessary to improve the ficiency of edge detection method, In other words Gaussian filtering helps reduction of detecting false dees. Next step isto calculate the magnitude and gradient ofthe edges in the smoothed image This is accomplished by filtering the smoothed image with a Sobet kernel In vertical and horizontal directions ‘The complete scan of image is done after receiving gradient magnitude and direction, to remove any @) 201;5(9;188200 unwanted pitels, which may’ not establish the edge. In {his step, hist Tocal maxima must be considered as edges ‘rough applying non-masimum suppression, Non- maximum suppression exchanges the smoothed edges i the frame ofthe gradient magnitudes to sharp edges. This sep is necessary to Keep every local matirourn in the fratient image, and remove any other detected eiges ‘hich are possibly false detections ‘The final step of canny edge detector algorithm is lysteresis thresholding. In this step, two threshold values are selected. The edges with intensity gradient treater than the maximum threshold are labeled as nire-edge" Similarly the edges with intensity gradient smaller than the minimum threshold are abeled as"non- fee’. Other edges between these threshold values are labeled as “sure-edge” i they are connected to another ‘re-edge, otherwise labeled as “non-edge”, Obviously, this step removes stall edges in the images that are possibly false detections, These threshold values are the ‘nly parameters ofthe method and are selected at 100 and 200 inthis work ‘Typically, the edge lines at the output of Canny method are thin lines. Therefore, clation is applied to binary edge image as morphological operation to thicken theline of objects om image. Because thin line may cause errors when finding a landing zone. The Binary image obtained after thi operation is named as objects image ‘The outputs ofthese steps are given in Fle 3 o) o Figure 3 steps of object detection. (a) FOV image, (2) Edge detection of FOV image, (<) Morphological operations are applied to binary image (objets image) 2.2.2. Searching fora landing zone (nce the objects inthe FOV image are detected, the next step Isto find the closest avalable zone fr landing ‘The first parameter to consider for ths step ls the ‘minimum ‘dimensions of the suitable landing. zone Obviously, this parameter depends on the altitude at hich the FOV mage Is taken As lustrated in Fig. 4, i fonder to find a relation between the dimensions and the altude, a calibration procedure is carried out on the images taken ative diferent altitudes, When thealtitude is 1 meter, the ideal dimensions are 60x60 pixels; onthe other hand, when the altitude is 5 meters the ideal

You might also like