Abstract

Video Surveillance and their installations are gradually being used to public services and association in order to obtain high level of security. In Video Surveillance applications, real-world Closed Circuit Television (CCTV) footage frequently creates new difficulties to object tracking because of to Pan-Tilt-Zoom operations, low quality of camera and different operational environments. The majority of significant difficulties are moving background, movement blur and rigorous size changes. However in the machine learning based video tracking system users have achieved better performance and exactness of object motion detection when compared to conventional video tracking systems. Particularly Convolutional Neural Networks (CNNs) have attains enhanced performance in object detection and it is being used to follow a more capable object tracking scheme. However CNNs based object tracking scheme becomes very challenging difficult to trace the object for less quality images. To enhance the quality and clarity of CCTV scenes, Motion Adaptive Gaussian Filtering (MAGF) denoising filtering is proposed in this paper, it is applied to three following frames, noise frames and detects the video movement region and still region. MAGF is performed based on the variation among the before and after video frame whether it is on movement or fixed. Within the MAGF, the Temporal Filter will be applied to stationary part of the video frame and Spatial Filter will be applied particularly to the movement part of the video frame in CCTV scenes. By using these filters noises in the video frames are removed and clarity of the video frames is enhanced to CCTV scenes. Then Kernel Support Vector Machine (KSVM) is proposed for multi-object tracking scheme and it is being utilized to follow a more capable object tracking scheme. In this work, make use of heterogeneous training video frames and video frames augmentation is investigated to enhance their multi-object tracking detection rate in CCTV scenes. Furthermore, it is proposed KSVM to make use of the objects spatial transformation parameters which calculate the development of intrinsic camera parameters and consequently adjust the object detector for higher performance.

Paper Details