Object Detection Using Haar Feature Selection Optimization

Object detection in still images is one of the common problems which is needed to be solved in a robust and reliable manner. Main focus on this work is the designing of classifiers based on Haar like simple features to obtain a good and efficient detection performance. This problem corresponds to the so called feature selection problem which is common in the pattern classifier systems. Classifiers used to detect objects are based on the simple Haar like features and these features are selected using systematic and general evolutionary based algorithm. The objective is to build a set of classifiers which respond stronger to the features present in object patterns than to non-object patterns, thereby improving the class discrimination between these two classes. This approach combines the classifier design with feature selection by using a genetic algorithm (GA). In the feature selection part of the algorithm a GA algorithm which the Haar features are encoded using their parameters in a single chromosome and optimized using genetic operators. During optimization the features which show similar characteristics in the parameter space are selected using a cluster based partitioning algorithm and thereby redundancy in the features is eliminated and a more compact Haar feature set can be obtained. Performances of the resulting chromosomes are measured using a fitness measure which is based on the separation of the two classes samples over a validation set. The resulting object detection structure is tested for near frontal face images in the cluttered background images