3D object identification based on global shape descriptors

The rapid development of laser scanning technologies leads to the emergence of new challenges and opportunities for the use of these technologies in various applications and fields. Typical examples of such tasks are classification and recognition (identification) of objects. One way to solve such problems is to use global descriptors of object shape, but it is not always possible to achieve the required accuracy with given requirements for the speed of algorithm. Difficulties arise when the base of reference objects in which search is carried out is big, but search algorithm is required to work in real time. The paper proposes a method of sequential application of global descriptors, allowing the first stage to produce a "rough" screening of obviously different objects, and then to apply more accurate algorithms on a significantly reduced object base. Let's formulate the proposed method of quick search of the object in the database. 1. A “rough” comparison of the input object with all database objects using the D1 descriptor. Remove/exclude from consideration of those objects of the database, the distance to which turned out to be greater than the established threshold β1 (“dissimilar” to the criterion D1 objects). 2. Then we calculate the eigenvalues λ1, λ2, λ3 of the covariance matrix, of the coordinates of the set of points Pi of the input ob- 3 ject. Exclude from consideration database objects, tor which ∑|λ^k_i -λ_i∣ > β₂, k - object index(number) in the database , β₂ - some i=1 threshold level. 3. Search the remaining set of objects using a more accurate descriptor (GASD, SPIN). Note that in order to improve the accuracy of recognition, it is possible to use not one specific descriptor, but several (GASD, SPIN, D1). We suggest sorting all database objects by the distance to the object you want to recognize and each database object is assigned a certain rank (its number in an ordered array). The number of arrays is determined by the number of descriptors used, and the total rank of each object is calculated as the arithmetic mean of its ranks in each array. Then the closest "similar" database object is the one that has the minimum rank. The effect reached at the same time consists in significant increase in speed of identification of an object (in certain cases much and more) without compromising accuracy. To improve the accuracy of object identification, the paper also proposes a method of simultaneous use of existing global shape descriptors. Experimental studies have shown that the application of the proposed method provides greater accuracy of object identification than any of the descriptors individually.

Keywords

Authors

References