الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 123 |  |  |
| | | from | 123 | | |
Abstract
This thesis proposes a real-time automatic people crowd density estimation method for overcoming the non-linearity problem, working with different densities and scales, enhancing the prediction error, and reduction of the computational time.
To cover most of the properties of the crowded scene, a newly used combination of features is proposed that includes segmented region properties, texture, edge, and keypoints. Edge strength is a suggested for use.
To select the best regression model for use in this work, the complete feature vector is utilized to train five different models (Gaussian process regression (GPR), random forest (RF), random projection forest (RPF), least absolute shrinkage and selection operator (LASSO), and K-nearest neighbours (KNN)). GPR and RF are the best ones. The cross validation technique is used to generalize the method in order to decrease the prediction error by selecting training and testing samples that cover all different cases.
Due to the existence of irrelevant, redundant, and correlated features, the prediction error increases. To remove such bad features, multiple feature reduction techniques is applied (filter, wrapper, and embedded). Using smaller feature vector gives the way to increase model capacity by using more number of training samples.
To avoid extracting features from all segmented regions within each frame, feature tracking is applied. Each region inside one frame is compared to all regions in the next frame. Features are reused rather than extracted if two regions matched.
The proposed method is tested over three different crowd counting datasets (UCSD, UCF, and Marathon). Based on the mean absolute error (MAE), the results show that the proposed work with the complete feature set is able to overcome the non-linearity problem and gives less error compared to previous works.
The best regression models across all experimental results are GPR and RF. LASSO gives the best reduced feature set that eliminates bad features. By applying feature tracking, the computational time reduced severely while maintaining the prediction error.