Activity Detection and Tracking in Video Surveillance

Abstract

Our research on following observations. Like activity detection and tracking. Motion detection algorithms like frame differencing and feature correlation. Using real-world video data to automatically isolate parking area suspects, which helps security surveillance. There has been a great deal of research in computers over the last few decades to detect objects in pictures and to track them by video. We developed an automated thermal video surveillance system to track pedestrians and detect situations where people may be at risk. Thermal imaging systems include human detection, human tracking and human activity identification. Extraction of moving objects by Thermal Digital Inputs has been improved. Video surveillance system can be use for real-time application. The classification approach is used to detect various suspicious activities such as looting, fainting, unauthorized entry, etc. Suspicious activities are defined using semantic approach and object detection is done using background subtraction. The objects are then classified as living or non living .

These objects detect by using correlation technique. In the end features & temporal information the events are classified as normal or suspicious. As the mathematical complexity is less and the efficiency of the approach is higher. Surveillance helps to protect them by donating close observation and monitoring of behavior. Video surveillance with intelligence is now use to identify or decision making according to the scenarios. IVS (Intelligent video surveillance system) based on image recognition is widely employed to effectively avert crimes and provide public security. High complexity of processing real time data and understanding of image contents, no application developed under cheap cost. To understand the suspicious activity detection methodologies used for detecting abnormal human behavior, tracing abandoned object, or unattended baggage etc., which led to an extensive comparison between various proposed methods. The outcome of the review is presented in form of various findings, which includes techniques and methods used to solve particular research problem, along with their strengths and weaknesses and the scope for the future work in area.

Introduction

Activity detection and tracking using real- time for analyzing out- door, far-field activities. Improving activity analysis with the help of intelligent control and failover mechanisms, built on top of low-level motion detection algorithms such as frame drafts and feature opacity. Tracking and recovering from both non-destructive errors and catastrophic errors to improve overall robustness and accuracy. Includes a two-tiered, hierarchical activity representation scheme for speculation tracking, and bottom-level data fusion and top-down information guidance.

To identify activity, we recommend effective representation of human activity based on tracked trajectories. We made a scheme that Prominent sample of interaction between group of people by identifying unique signatures in prominent locations and speed participants and their respective positions and speeds. We demonstrate our techniques using real-world video data to detect and recognize behaviors in a parking lot set- ting. In particular, we are able to distinguish routine follow-up behavior from suspicious and potentially dangerous, stinging behaviors, which can help with security surveillance.

By given knowledge to object recognition we can detect object in an image. As human brain has ability to earning new objects and recognition them. We observe problems of object recognition like human motion, how human execute their habits. This differs from other human detecting techniques like how human shape, color and texture. We can use applications of human-computer interaction, user interface design, robot learning, and surveillance. For Night Vision and tracking we can use Thermal Imaging based Video Surveillance systems. Observing pedestrian traffic areas and detect danger automatically becoming important. Therefore, there are rich chances to develop automated intelligent vision-based monitoring systems which can secure human user in the process of risk detection and analysis. Tracking is the fundamental component for surveillance applications. Before recognition the pedestrian must be tracked first.

Best option to be used privately and as publically to protect people is the Video Surveillance. To detect any abnormal /suspicious activities efficiently Video surveillance is the best option. Currently there are many surveillance systems operated by humans. So they require continuous human attention to detect any abnormal activity. Semantic based approach is used to define & detect the suspicious activities. The framework of the system consists of defining suspicious activity, background subtraction, objects detection, tracking & classification of activities. The suspicious activities are defined using Semantic approach which applies the human understanding of the activity. The motion features between the two/different objects are extracted to detect the behavior. The disadvantages of the machine learning such as unavailability of standard datasets, generalizing ability of the classifier can be overcome using the semantic based approach. The content of the paper are as follows- In section two the work done by different researchers in this field is studied. Section three describes the system flow and the working of the system. In section four the results which are obtained after experiment are shown. The last section is the conclusion which is drawn from the results obtained.

Recognition of image, as Computer Vision to be defined as ‘the feature in a digital image or video and process of identifying’. In recent years Intelligent techniques have been . Image recognition and computer vision has emerged out to be an Intelligent Video Surveillance System. Generally maintaining records of monitoring scenes replacing human eyes by Conventional video surveillance system. In unusual situation videos can be analyzed . This method reduces the real –time security and utility system. Without any human involvement Such systems can make intelligent analysis for real time video image sequence. By detection, classification, tracking, and behavioral analysis It can also monitor scene.

Methodology

Activity Detection and Tracking When people are away, we could approximately describe every person as a “blob” and using single object state describing the trajectory as a function of time. The goal of activity detection and tracking is then to optimally infer the state vectors for people in the scene from multiple cameras x(~) = [p(‘)(t),p(i)(t),p(i)(t)]T,O 5 i < m, where misthenum- her of cameras used, and fuse such 2D state estimates from multiple sensors to derive a consistent, global 3D estimate.

Real-time, control and fail-over mechanism-on top of correlation-based tracking-to deal with noise and low level frame differencing , short periods of absence , scene clutter , merging of silhouettes, and long periods of occlusion of activities of a camera view. Formulation based on powerful hypothesis and verification, which has been alternatively named as Monte Carlo filtering , particle filtering 191, genetic algorithms 141, condensation and I condensation. Mathematically, all state estimation algorithms are geared toward estimating the following conditional probability in an iterative manner: x = [P(t),P(t),P(t)]T. I

P(x;~~xo, ZI~ZZ, 23 .’. zn) = P(x,+~xI,zz,z~. ..,zn) P(X,+lXZ> 23,. .. , Zn) (1) = = … = ~(x2l~n-1,zn) o( p(zn~x,)P(x,~xn-l)P(~n-l~ where sensor data are denoted as z and states as x (x- and xf denote, respectively, the state before and after the sensor measurement made at a particular time is incorporated.) The difference is in the forms the state prior and the various noise processes inherent in sensor could assume measurement and complexity of propagation process. The utility of a general hypothesis-verification formulation, over traditional linear state estimation algorithms such as Kalman filtering, is that the noise processes do not have to he Gaussian and state propagation shouldn’t be unimodal. Allowing multiple competing hypotheses maintaining and contributing to the state estimation.[image: ] Different incarnations of the same hypothesis- verification principle all comprise the following essential component^:^  A collection of candidate states and their likelihood estimates that are initialized, propagated, and updated over time,  a state propagation mechanism a state verification mechanism using sensor feedback and a re-normalization process to refresh and/or regenerate the collection of candidate states and their associated likelihood @(xklx., 21,. . . ,z,)). Employing region growing and frame differencing in initial detection of activities. Moving region identified By camera image, multiple hypotheses are postulated. Single person hypothesis representing region as a moving person, characterized by state comprising the position, velocity, and acceleration of the moving region. Hence, more than one states are maintained which track these hypotheses. Using instantaneous velocity with in time the position is propagated and acceleration recorded in the state vector. Such alternate hypotheses could based on object’s signature discovered by search in reduced-resolution image over a wider image area, or the difference in exact mechanism for realizing . For ex- ample. standard condensation algorithms genemte and update candidate states based smctly on the Bayesian formula. while imporonce-based condensation allows auxiliary information (e.g., from senson) to be brought in to better position candidate states far efficiently exploring the state space.

Activity Representation and Recognition To recognize individual and gathering activities and interactions directly on staying properties using the recovered trajectories, without elaborating pars of state vectors against pre-established Markov models. For example, loitering movements are known by large variance in direction, but small variance in position.