الفهرس | Only 14 pages are availabe for public view |
Abstract Pedestrian detection and tracking play a crucial role in security and intelligent surveillance, robotics, and vehicles. Occlusion handling is a challenging worry in tracking multiple people. Adaptive and advanced solutions are required to track pedestrians for video surveillance purposes accurately. This thesis presents a method of tracking multiple people under some occlusion conditions. In this study, we proposed a system that combination of detection and Object Tracking algorithms, we take advantage of the high efficiency and celerity of Retina-Net detector algorithms with Kernelized Correlation Filter (KCF) for tracking and handling the occlusion of multiple pedestrians in real-time performance. We applied the transfer learning concept by using the Res-Net101 deep networks, which is trained before on common object context (CoCo) dataset to extract the feature map relying on trained weights to reduce the time of training and fine-tuning hyper-parameter to achieve the best result in detection. The framework was tested on Caltech pedestrian database. The pedestrian detection accuracy was measured by using F1-score that is achieved 81.5%. Also, the proposed system achieved Miss Rate (MR) equals 28.8 %, precision= 95.3 %, and recall equals 71.1% The achieved results show the promise of our proposed technique to detect and tracking multiple pedestrians in a video scene. |