Automated landmine detection by means of a mobile robot

Abstract

Millions of antipersonnel landmines are left in the ground after past war conicts across many countries. Being functional for more than 50 years they provide a lot of humanitarian and economical problems long after the conict is _nished. Cleaning the existing mine_elds, called humanitarian demining, is required in order to return the large areas of the land to normal use and save the local civilians from the danger. Currently, the only fully trustable solution for this problem is the manual clearance which is itself a very dangerous and slow procedure. Automation of the humanitarian demining may change the situation providing a faster approach which eliminates the participation of humans on the mine_eld. This work is a part of the e_ort toward the development of such approach. Automation of the humanitarian demining meets a lot of technical problems which currently do not have e_ective solutions. This work covers the ones related to the automatic detection of antipersonnel landmines assisted by a mobile scanning platform which carries the landmine detection sensors. The landmine detection approach developed in this work assumes the employing of several nonselective sensors most widely used for landmine detection which include metal detectors, infrared sensors, and ground penetrating radar. The approach has a multi-stage structure and is based on feature-level sensor fusion strategy. This process is understood as a step-by-step reduction of the false alarm rate depending on the quality of the available sensor data. During the _rst stage the sensor data are processed in order to distinguish all objects suspected to be landmines against the background. For this purpose a novel online algorithm was developed. It allows to detect the object during the robot movement and is hardly sensitive to the quality of the sensor data. The consequent stages are performed in order to recognize the landmines among the detected suspicious objects. A number of new classi_cation features were developed in order to perform this recognition. Based on the feature analysis a concept of selective training specially suited for the landmine recognition task was developed. This technique allows to account for the high overlapping of the classes and multimodal distributions of the classi_cation features. Finally, a concept of dominant class was introduced in order to provide high levels of detection rates even in case of poorly separated classes. Being specially designed for the speci_cs of landmine detection the proposed algorithms allow to improve the results. In order to assist the gathering of the sensor data, the problems related to the effective sensor data gathering,path planning and localization of the platform are also addressed. The developed solutions are implemented on the previously created pneumatic scanning platform acting as a prototype demining robot. A number of practical solutions improving the platform localization were developed. The positioning of the robot is based on its odometry, compass and a novel vision system which are combined together by means of a Kalman _lter. The vision system employees a simple CCD camera and is guided by a novel algorithm for the detection and association of natural landmarks found on the ground surface. Finally, considering the landmine detection problem in the scale of mine_eld the problem of the _eld exploration is addressed. Assuming a general case in which the mine_eld may be populated with some obstacles in unknown positions an algorithm for online unknown area coverage was developed. The algorithm guarantees regularity of the robot path necessary for the mapping of sensor data and the safety of the robot by planning its path only inside already covered area. The developed algorithms were implemented in a form of control software for the real platform. Testing of the proposed ideas in simulation and in real conditions (on a test mine_eld) provided promising results showing the perspective of the developed concepts. Based on the experimental results the recommendations for future work are formulated. The automatic landmine detection task raises a number of challenging problems which have connections to other areas of robotics, pattern recognition and control. In this regards the development of the methods proposed in this work was considered in a more broad sense. Thus, the results of this work can be used in the adjacent _elds of robotics: automatic subsurface sensing with online reaction to the found target, pattern recognition in case of poorly distinguished classes, and online unknown area coverage required for cleaning, grass cutting, agriculture, etc.