Unmanned Aerial Vehicle Emergency Landing Site Identification System Using Machine Vision Sumair Aziz, Rao Muhammad Faheem, Mudassar Bashir, Adnan Khalid, and Amanullah Yasin Center for Advanced Studies in Engineering (CASE), Islamabad, Pakistan Email: {sumair.aziz, adnan_uet38}@yahoo.com, {rao.m.faheem, mbuetect322}@gmail.com, amanyasin@case.edu.pk Abstract—Unmanned Aerial Vehicles (UAVs) rely on navigation commands from autonomous flight control system or from Ground Control System (GCS) via line-of- sight wireless data link. UAV needs to perform immediate landing on predefined airfield in case of extreme emergency like navigation, data link, engine or control surface failure, that cannot be accomplished in some cases and accidents can occur which can result collateral damage as well. This paper presents the system design which can discover the appropriate area within the surroundings for immediate landing in case of emergency. The proposed system design consists of two stages. During first stage, system takes top view images from UAV onboard camera, then image processing algorithm extracts and refine the attributes of the image. In second stage, machine learning based algorithm evaluates the results from previous stage, and based on its previous training, decides whether the area visible in image is good for safe landing or not. We implement proposed system design in MATLAB and the approach used is validated with experimental results on test data. Proposed system design uses combination of simple techniques, which makes it less computationally intensive, having reduced latency, low implementation cost and easy to implement on high speed real time hardware like FPGA/ASIC. Index Terms—vision based landing, emergency landing, machine vision, machine learning I. INTRODUCTION Unmanned Aerial Systems are extensively used in commercial and military sectors e.g. security & control, aerial reconnaissance, aerial policemen and crowd monitoring, security watch, maritime search and rescue, oil and gas pipeline monitoring, disaster effects management, rescue and clear up effort supervision, disaster damage estimation, crop management, telecom relay and signal coverage survey, oil and gas exploration and geophysical surveys. UAVs accident rate is 100 times more than manned aircraft. Although UAV failure rate is about one per 2000 flight hours but analysis shows that for every 1000 flight hours there is at least one mishap [1]. UAVs mostly crash because of mechanical failures and data link loss. In case of data link loss UAVs are programmed to fly in a circular pattern until the links are restored. In worst-case scenarios, they are supposed and programmed to return autonomously to their launch base Manuscript received April 20, 2015; revised November 2, 2015. on preprogrammed route information using GPS navigation. In more than a quarter of the accidents investigation had revealed that data links were lost and UAVs had not returned to the base. Satellite/data link connections can be lost when a UAV take sharp turns or drop altitude too quickly due to failure of engine or control surfaces. Link losses can also occur in case of Data Link Station failure e.g. power loss or equipment malfunction. This link can also be disrupted due to uneven terrain and mountains between GCS and UAV due to directionality of microwaves. 45-50% of total crashes were due to engine and mechanical failure. That could be due to oil leakage, propeller malfunctioning, control surfaces (flaps, ailerons, rudder) failure [2]. In the past UAV designers had not carried out much research in improving the emergency landing site selection system. In last decade some valuable research had been carried out in the field of computer vision and machine learning to develop the emergency landing systems. The family of researchers which developed the emergency landing systems using machine learning have used Support Vector Machine (SVM) and Artificial Neural Network (ANN) algorithms along with digital image processing techniques to select the appropriate site for landing [3], [4]. Since SVM is very complex and computationally intensive technique, and ANN requires large data for training as a result it requires more training time, so both these techniques cannot satisfy the fast changing demands in emergency or forced landing area selection systems. A better and computationally less intensive approach of classification i.e. K-Nearest Neighbor (KNN) technique can be used to get the desired results. Alongside machine learning algorithms, image processing techniques like texture/features extraction are also required to get the useful information in the images data for classification purposes. Many researches have used the texture features in order to classify the images. Image data is very large, and it becomes very complex and time taking for the machine learning classifiers to train and then predict. The proposed approach used dimensionality reduction concept, in which few most important attributes are extracted from the large image data. Then extracted attributes are passed to machine learning algorithm for training and then real time test cases. Dimensionality reduction gives excellent improvement of performance in terms of processing speed and complexity. Journal of Image and Graphics, Vol. 4, No. 1, June 2016 ©2016 Journal of Image and Graphics 36 doi: 10.18178/joig.4.1.36-41