Transcript Document
Sound-based Techniques for Navigation Josh Markwordt October, 24 2007 COMP 790-058 Overview • Basics of Sound • Active Sonar • Passive Sonar 10/24/2007 2 Basics of Sound • We perform motion planning and navigation everyday • In addition to sight, we use our other senses including sound to aid our decisions making • Our audio sense differs in a number of beneficial ways from our sense of sight: – Sound is omni-directional – Sound can move around obstacles that completely obscure line of sight – Sound can be a good way to quickly estimate the general direction of something 10/24/2007 3 What is Sound? • Sound is a mechanical disturbance that propagates through air as a wave • Longitudinal wave, meaning motion of particles is along the direction of propagation • Compressions and rarefactions in air pressure (or other medium the sound is being transferred in) 10/24/2007 4 Properties of Sound • Wavelength () is measured from crest to crest • Frequency (f) refers to how many cycles pass by per second at a given point • Phase (): Measures the progression of pressure at a point between a crest and a trough 10/24/2007 5 Diffraction • Sounds with wavelengths similar to the dimensions of obstacles diffract around those obstacles • High frequency sounds, with short wavelengths, do not diffract around most objects but are reflected or absorbed • A sound shadow is created behind objects, which is perceived as a decrease in the loudness of a sound 10/24/2007 6 Interference • Resultant pressure is the linear supposition of the two signals • If they are in phase, the signal will magnify the amplitude of the wave resulting in “constructive interference” • Signals that are out of phase produce “destructive interference” 10/24/2007 7 Overview • Basics of Sound • Active Sonar • Passive Sonar 10/24/2007 8 Active Sonar • Creates a pulse of sound from a transmitter and listens for reflections of the pulse • Time from transmission of a pulse to reception is measured to determine distance • An array of sensors can be used to gain more accurate information 10/24/2007 9 Probabilistic Occupancy Grid • Number of difficulties with processing sonar data, mainly due to the sensors themselves – Low angular resolution – Errors due to multiple reflections or specular reflections away from the sensor • This leads to a probabilistic approach to the recovery of spatial information from sonar range data 10/24/2007 10 Probabilistic Occupancy Grid • Sonar reading R corresponds to a range value r returned by a sensor • Define a conditional probability function P(cell C is occupied | sensor reading R) • Informally measures our confidence that cells inside the cone of the beam are empty 10/24/2007 11 Probabilistic Occupancy Grid • Occupancy grids are 2D or 3D tesselations of space into cells • Each cell has probabilistic estimate of its occupancy • Each cell C has a state variable s(C ), a discrete random variable with two states OCCUPIED and EMPTY such that P(s(C ) OCCUPIED) P(s(C ) EMPTY) 1 10/24/2007 12 Bayesian Model for Updating • Can avoid combinatorial explosion by assuming – Each cell in the grid is independent – The superposition of many readings will eventually converge to the correct state of each cell • Bayes’ Theorem tells how to update or revise probabilities in light of new evidence 10/24/2007 13 Bayes’ Theorem P( si ) P(e | si ) P( si | e) P ( s j ) P (e | s j ) • • • • • j Where s i is one of n disjoint states being estimated e is the relevant evidence P( si ) is the a priori probability of the system being in state i P(e | si ) is the probability that evidence e would be present given that the system is in state P(si | e) is what is needed for decision making, namely conditional probability that the system is in state i in light of the evidence e 10/24/2007 14 Bayesian Model for Updating • In our system, the evidence is sensor range readings R • Desired probabilities P(OCC |R) and P(EMP|R) P(OCC | R) P(OCC) P( R | OCC) P(OCC) P( R | OCC) P( EMP) P( R | EMP) • Since P(R|EMP) = 1 – P(R | OCC) we can write it just in terms of one of the probabilities P(OCC | R) 10/24/2007 P(OCC) P( R | OCC) P(OCC) P( R | OCC) (1 P(OCC))(1 P( R | OCC)) 15 Bayesian Model for Updating • This formula has several useful properties – It is commutative and associative (data can be incorporated in any order) – Combining evidence E with the prior probability for UNKNOWN (P(OCC) = 0.5), gives E as a result – Conflicting measurements of the same strength cancel (OCCUPIED +EMPTY = UNKNOWN) • Initialize map by assigning equal probability of P(OCC)=P(EMP)=0.5 (or UNKNOWN) to each cell 10/24/2007 16 Probabilistic Occupancy Grid • Graphic corresponds to reading taken by a sensor positioned at the upper left pointing to the lower right • The plane shows the UNKOWN level • Values above the plane represent OCCUPIED probabilities • Values below represent EMPTY probabilities 10/24/2007 17 Sonar Map Example 10/24/2007 18 Active Sonar Maps • A single set of sonar readings – provides substantial information about empty space, but is weak in recovering object shape – Immediate motion decisions are possible 10/24/2007 19 Active Sonar Maps • Only through the composition of several views is more spatial structure recovered • Wide angle of beam sometimes precludes the detection of free space between close objects • Phantom obstacles can be generated due to multiple reflections • But over several sets of readings sonar is able to recover good descriptions of surfaces 10/24/2007 20 Active Sonar Maps • Resulting active sonar maps are very useful for navigation – Much denser than ones generated by previous sparse stereo vision programs – Computationally about an order of magnitude faster to produce • Autonomous navigation system that uses an A*-based path planner to obtain routes in these maps 10/24/2007 21 Overview • Basics of Sound • Active Sonar • Passive Sonar 10/24/2007 22 Passive Sonar • Passive sonar listens without transmitting • Must rely on the sound of the object being detected • Advantages – Power consumption of passive sonar is very low – Data can be used for target identification or to pinpoint location of the emitting sound source • Disadvantages – Can be more difficult based on the amount of noise or stealth of the object – Need to identify an object by sound 10/24/2007 23 Beamforming • Takes advantage of the distance from an object to each receiver being different • Each recording will be a phase-shifted replica of the others • Shift the signals in the opposite direction in order to get them in phase • Results in constructive interference to amplify the desired signal 10/24/2007 24 Beamforming 10/24/2007 25 Beamforming Pseudocode 10/24/2007 26 Passive Sonar System • EvBot II equipped with eight microphones • Tested with various real-object sounds such as helicopters, trucks and airplanes 10/24/2007 27 Passive Sonar System • Able to effectively track a moving object emitting a sound 10/24/2007 28 Conclusion • Sonar maps are useful for localization in motion planning tasks • In combination with stereo vision can provide a more complete view of the world to autonomous robots • Passive sonar can be used to track an object, but it’s not as practical as active sonar • Questions? 10/24/2007 29 References • Sensor integration for robot navigation: Combining sonar and stereo range data in a grid-based representation Alberto Elfes, Larry Matthies 26th IEEE Conference on Decision and Control, Vol 26 1987. • Beamforming: A versatile approach to spatial filtering Van Veen, B.D., Buckley, K.M., IEEE ASSP Magazine Vol 5, Issue 2 1988. • Passive sonar applications: target tracking and navigation of an autonomous robot Mattos, L., Grant, E. IEEE International Conference on Robotics and Automation Proceedings, Vol 5, Issue 26 2004. • A sonar-based mapping and navigation system Elfes, A. IEEE International Conference on Robotics and Automation Proceedings, Vol 3 1986. 10/24/2007 30