Information Technology Reference
In-Depth Information
Combined feature vectors, stemmed from descriptor evaluation measures (Table
12.11 ), turned out to outperform single feature based vectors. In case of PERSON
dataset, VectorPerson achieved the best results (on average), while Vector
Vehicle provided the worst outcome. For VEHICLE dataset, VectorVehicle
feature vector results are at the top (excluding local image features that per-
form the best and cannot be mixed with other descriptors), however four single
descriptors ( HistFull , MomentInvGPSO , Hist and LBPHist ) provide some
slightly better results; VectorPerson feature vector is penultimate. The vector
VectorBoth proposed as the universal one, is characterized with a mediocre per-
formance; it is better than the combined feature vector not designed for the object
type, but significantly worse than the suited combined feature vector and many single
descriptors. This means that it is hard to provide the content of a feature vector that
would be universal and accurate of all types of objects; the feature vector should be
adjusted to the object type in order to obtain a high performance.
12.8 Conclusions
The problem of object re-identification in multiple cameras required solving the
problems of extracting important features of moving objects from a video stream
and efficient comparing the descriptor sets in order to match the same object in two
cameras. In order to select only distinctive descriptors, a comparative analysis of
many visual features was presented. A wide range of descriptors have been studied,
including the ones based on colour, texture, gradient and local image features. Eval-
uation of the descriptors was performed for two image datasets containing persons
and vehicles. On the basis of the experiments, features that are best suited for object
re-identification in multi-camera environment were selected.
The proposed method of descriptors evaluation is based on the measurement
of direct dissimilarity between pairs of images of objects. This method is supple-
mented with two other methods known from the literature. In order to combine the
results achieved for various datasets, four aggregation measures were introduced.
Their goal is to find descriptors that provide the best results based on the relative
ranking, whereas they are simultaneously characterized with large stability, i.e. their
effectiveness does not depend on the selection of particular objects in the dataset. The
proposed descriptors are evaluated practically with object re-identification experi-
ments involving four classifiers to detect the same object after its transition between
cameras' fields of view.
Results achieved show that there are no visual descriptors that provide the
best results for both datasets, simultaneously. In case of human images, the best
results were provided by descriptors based on colour, gradient, and image statis-
tical moments. In case of vehicles, the favourable ones are descriptors utilizing
colour information, image statistical moments, and local features. The most uni-
versal descriptors, providing adequate effectiveness both for humans and vehicles,
are the statistical-based ones.
 
Search WWH ::




Custom Search