Object Detection using Dimensionality Reduction on Image Descriptors

The aim of object detection is to recognize objects in a visual scene. Performing reliable object detection is becoming increasingly important in the fields of computer vision and robotics. Various applications of object detection include video surveillance, traffic monitoring, digital libraries, navigation, human computer interaction, etc. The challenges involved with detecting real world objects include the multitude of colors, textures, sizes, and cluttered or complex backgrounds making objects difficult to detect. This thesis contributes to the exploration of various dimensionality reduction techniques on descriptors for establishing an object detection system that achieves the best trade-offs between performance and speed. Histogram of Oriented Gradients (HOG) and other histogram-based descriptors were used as an input to a Support Vector Machine (SVM) classifier to achieve good classification performance. Binary descriptors were considered as a computationally efficient alternative to HOG. It was determined that single local binary descriptors in combination with Support Vector Machine (SVM) classifier don\u27t work as well as histograms of features for object detection. Thus, histogram of binary descriptors features were explored as a viable alternative and the results were found to be comparable to those of the popular Histogram of Oriented Gradients descriptor. Histogram-based descriptors can be high dimensional and working with large amounts of data can be computationally expensive and slow. Thus, various dimensionality reduction techniques were considered, such as principal component analysis (PCA), which is the most widely used technique, random projections, which is data independent and fast to compute, unsupervised locality preserving projections (LPP), and supervised locality preserving projections (SLPP), which incorporate non-linear reduction techniques. The classification system was tested on eye detection as well as different object classes. The eye database was created using BioID and FERET databases. Additionally, the CalTech-101 data set, which has 101 object categories, was used to evaluate the system. The results showed that the reduced-dimensionality descriptors based on SLPP gave improved classification performance with fewer computations