الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 138 |  |  |
| | | from | 138 | | |
Abstract
In recent years, many researchers focused on human activity recognition systems according to their important role in preventing crimes before occurrence of danger- ous. The need for advanced surveillance systems became urgent to overcome acts of sabotage and to secure persons and buildings. This thesis presents a video based system for recognizing different continuous human activities in real time using a single stationary camera. The main idea of this system is to detect and extract the features of moving objects in each frame and associate the detections of the same object over time. The first feature part is constructed by extracting the object con- tour and estimate its normalized Fourier descriptors. The other part is extracted by the shape moments. We combined the shape moments to produce an invari- ant feature that is invariant to rotation, translation and scaling. We adopt the multi-class support vector machines (MSVM), Naive Bayes and neural networks classifiers. The MSVM shows better performance than the other methods of clas- sifications with a recognition rate up to 94.15%. We evaluated activity recognition on 325 videos of thirteen distinct human activities (e.g., Walking, Running, Jump- ing, Hand-waving, Bending and some suspicious activities like Kicking, Punching, Fall floor and Shooting gun, etc.) recorded for 260 different persons. Experimen- tal results on three data set Weizman, Kungliga Tekniska hogskolan (KTH)and Human Motion Database (HMDB) validate the proposed system reliability and efficiency. Also the human activity recognition has been considered using the Convolutional Neural Network. This is evident in the emergence of a number of convolutional neural network architectures such as LeNet-5, AlexNet and VGG16 and modern architectures such as ResNet, Inception V3, Inception-ResNet, Mo- bileNet V2, NASNet and PNASNet. The main characteristic of a convolutional neural network (CNN) is its ability to extract features automatically from input images, which facilitates the processes of activity recognition and classification. In addition, CNNs have achieved perfect classification on highly similar activities that were previously extremely difficult to classify. In this thesis, we evaluated mod- ern convolutional neural networks in terms of their human activity recognition accuracy, and we compared the results with handcrafted features (HCFs) repre- sented by statistical features. Previous experiments have shown that convolutional networks already derive more complex and related features with every additional layer. In this part of the thesis, we used two public data sets, HMDB (Shooting gun, kicking, falling to the floor, punching) and the Weizman dataset (walking, running, jumping, one hand waving, bending, jumping in place, two-hand waving, skipping, jumping jack) to evaluate our handcrafted features method and differ- ent modern CNN architectures and to compare the performance with the latest methods. Experimental results indicated that the CNN with NASNet architectur achieves better performance of the six CNN architectures on both human activity data sets (HMDB and Weizman) but when comparing the performance with hand- crafted features methods we found the superiority of handcrafted features with a support vector machine classifier especially on the HMDB data set. This is due to the differences between the videos in each data set: the Weizman data set videos contain a single person performing a single activity and have a fixed background, but the HMDB data set videos contain group activities, variable backgrounds and some objects are occluded. The locations of the objects are not available when they are occluded by other people or things; thus an approach such as using a Kalman filter is urgently needed to counteract these missing measurements.