Improved occupancy grid mapping in specular environment

Abstract

This paper addresses the improved method for sonar sensor modeling which reduces the specular reflection uncertainty in the occupancy grid. Such uncertainty reduction is often required in the occupancy grid mapping where the false sensory information can lead to poor performance. Here, a novel algorithm is proposed which is capable of discarding the unreliable sonar sensor information generated due to specular reflection. Further, the inconsistency estimation in sonar measurement has been evaluated and eliminated by fuzzy rules based model. To achieve the grid map with improved accuracy, the sonar information is further updated by using a Bayesian approach. In this paper the approach is experimented for the office environment and the model is used for grid mapping. The experimental results show 6.6% improvement in the global grid map and it is also found that the proposed approach is consuming nearly 16.5% less computation time as compared to the conventional approach of occupancy grid mapping for the indoor environments.

Introduction

Autonomous navigation in the field of mobile robots is a highly challenging area which deals with the complex mechanical and electronic design, control algorithms, artificial intelligence (AI), path planning and reasoning, etc. The robot has to deal with unstructured and highly dynamic environments where the objects like humans or other robots may be moving continuously. In such situations, confirmation about the uncertainty in the sensory information plays an important role in correct decision making for mobile robots. The occupancy grid mapping is considered as an effective method of map making [1], [2], [3] which was already used in the past, where the estimation of the probability of individual cells depends upon the output of the range sensor (sonar, vision, laser, IR, etc.). The uncertainty in sensory information is very common as a sensor cannot measure all attributes of perception during the navigation process.

This study deals with the handling of specular reflection in ultrasonic (sonar) sensors in which the energy from the device is reflected at an angle by the surface. Since the sensor reads time of flight, hence the reflected energy gives the wrong estimation after striking other multiple surfaces. It causes incorrect decision making in mobile robot navigation, particularly for indoor applications where objects like tables, walls or flat surfaces are not always perpendicular to the acoustic axis of the sonar sensor beam. To overcome such problems, a fuzzy logic algorithm is proposed which helps to discard the specular information. The proposed method is a refinement of sonar data used for occupancy grid mapping. The first part of the study deals with the detection and elimination of the specular information and the second part deals with the updating of the grid map using a Bayesian approach. The experiment is conducted for an indoor office environment and the performance of the conventional occupancy grid mapping is compared with the results of the occupancy grid generated after the implementation of the newly developed fuzzy technique. The qualitative comparison shows the improvement in the results representing the overall occupied and empty area is very near to the reference map. Quantitative comparison shows that the implementation of the proposed technique has brought about a 6.6% improvement in the occupancy grid. The experiment was conducted in three different environments and it is found that the proposed approach is consuming nearly 16.5% less time as compared to the conventional approach of mapping using a sonar sensor. Thus, the proposed method is quite attractive for real time applications.

A sonar (ultrasound) sensor is a proximity sensor which uses sound to measure the relative distance between the sensor and objects in the environment. Most robots today employ sonar sensors due to their low-cost and low-power consumption. Sound waves are emitted from a sensor and then detected back after they bounce back from the objects in the environment. The time difference between the emission of sound and detection of sound is used to determine the distance to an object. Sonar sensors are widely used in mobile robots for path planning, ranging and navigation [3] and used in grid map updating using Bayesian, evidential and fuzzy [4], [5], [6] algorithms. Sonar reliability depends upon the structural features of the obstacles present in common places [7]. Kuc [8] presented a method for discriminating planes, corners and edges. Others [9], [10] presented the application of a sonar ring in mobile robot navigation. The limitations of sonar (specular reflection, cross talking, etc.) have been considered as low efficiency parameters of sonar by many researchers [11].

Specular reflection occurs when the wave hits a surface at an acute angle and the wave bounces away from the transducer [12]. Ideally, all targets/objects should have a flat surface perpendicular to the transducer at the optic axis, but it rarely happens. In most of the cases, the reflected signal may get further reflected by a third object, and so on, until by coincidence it returns some energy back to the transducer. In that case, the time of flight will not correspond to the true relative range [13]. Even with severely acute angles, the surface is usually rough enough to send some amount of sound energy back. An exception to this is a glass surface that induces serious specular reflection, which is very common in hospitals and offices. Even so, if the surface roughness of the target is larger than the wavelength of the impinging beam, it will act like a point reflector and scatter energy equally in all directions, called diffused reflection of energy from the target. Kurt [14] distinguished the diffused and specular beam by using Bayes’ rule with re-evaluation of the probabilistic mixture of the two detected densities. Leonard and Durant-Whyte [15] pointed out that the strongest readings from the sonar are actually specular reflections. Drumheller [16] estimated the specular reflection from the sonar penetration conditions: the free space hypothesis of a sonar reading should not impinge on a high-confidence surface. Knowledge of the environment either a priori or acquired from previous sensory information can be used to estimate the probability of a given reading being specular. Jong [17] handled the specular reflection by employing two parameters, (a) range confidence factor (RCF) and (b) orientation probability. Zou Yi [18] noticed that the performance of RCF depends upon range parameters. The selection of these range parameters for individual observations slows down the evidence accumulation. Zou [18] proposed modifications but the nature of the environment has not been considered in his research. Kurt [14] used the MURIEL method to determine the probability of specularity for each individual range reading.

The MURIEL [14] algorithm is a simple method of computing the probability of specularity that is local to a given cell. Using this method the author explained that if a cell is occupied, then from some poses of a robot, the sensor will give a surface hypothesis reading for the cell. If a cell has enough such readings, it is assumed that it is (probably) occupied, and computes the specular probability for all free space readings at the cell, based on a measure of how strong the surface hypotheses are. This method has the advantage of being quickly computable, and yields a reasonable value for specular probability in many situations. But, it has the following drawbacks. First, it treats all free space readings as having the same specular probability value, based on surface readings at one cell only. Obviously, there may be some free space readings that are much more likely to be specular, because their free space readings impinge on other occupied cells. Second, different cells may conclude that a single reading is specular or diffused, e.g. if a cell has no surface readings, any free space readings are assumed to be diffuse and that cell is considered specular.

Although the examples were based on sonar sensors, the MURIEL method showed improvement in the performance of occupancy grids with any sensor that has a specular component, such as radar. Kurt also highlighted some required modifications of the MURIEL method, which have not been investigated. One is a global assessment of specularity, based on the sonar penetration condition: any new reading could be checked in this way. The author expected that it would be hard to keep track of older readings and re-evaluate them whenever a relevant cell is modified.

We propose a fuzzy logic based approach that eliminates the sensory information generated due to specular reflection. The proposed approach is keeping the record of the previous reading, i.e. considering only two successive increments. This reduces the computational burden that was required in other earlier methods.