Pssst… we can write an original essay just for you.
Any subject. Any type of essay.
We’ll even meet a 3-hour deadline.
121 writers online
Detection of smoke caused from public cigarette smoking has been a common goal of many systems since it has applications not only for health concerns but also because smoking is a prominent cause of fire in public areas. A review of work done previously for recognizing and detecting smoking activity in public areas is presented and various models and techniques that are available for training of a neural network that have applications in object recognition in real time are discussed. Sensor based system proposed by Y. Liu et al is a widely used passive smoke detection sensor. However such sensors that detect the smog from tobacco products are not sensitive enough for rooms with large area. Also it can be costly to install such a sensor in a room and may also require reorganization of the room to allow sensor installation. P. Wu et al suggests a vision based system that uses color based histogram to draw interaction between the cigarette and its human holder. Since this approach is based on identifying human gestures to indicate the smoking event, it is prone to falsely detecting an event where the gesture is similar to holding a cigarette such as a pen held by someone. Another approach used to detect smoking makes use of the pattern that the smoke leaves on WiFi signals as illustrated by Zheng, Xiaolong, et al. However this method gets restricted to indoor environment where there are WiFi signals available that will be impacted by the smoking pattern, especially in India where WiFi has not become ubiquitous.
Smoke detection is an actively researched topic because of the challenge it presents in identifying its visual characteristics in outdoor spaces where the environment is mostly unknown. Tian, Hongda, et al describes an image separation approach where the smoke component is distinguished from the background by assuming the image to be a linear blend of smoke and background. To obtain the smoke component, the author suggests using Gaussian Mixture Model (GMM) as a pre-processing step for foreground extraction. However, when the environment is highly dynamic such as a shopping mall, the background subtraction approach lacks significance due to multiple moving targets. To overcome the drawbacks of using smoke detectors, Çelik, Turgay et al makes use of the colour information along with fuzzy logic to determine presence of smoke. This approach, though is effective in detecting the heavy smoke caused during a fire, it can miss the detection of cigarette smoke which is light and transparent in colour with low density. Another technique that is used in smoke detection captures the motion of blobs of smoke in each video frame and then applies a connected component analysis for the entire video. This method is applied by Brovko, N. et al using optical flow algorithm to detect the motion and estimate if it lies within 0-45° to be characterized as smoke and by Filonenko, Alexander. In the latter approach, the author boosted the performance of the algorithm by employing parallel processing of CUDA GPUs to process both low and high resolution videos.
Deep learning is a powerful computational model composed of multiple layers that is capable of learning the data with high level of abstraction. With the evolution in deep learning, the need for hand tuned machine learning used previously has reduced significantly. It has also led to the development of state-of-the-art systems for object detection, speech recognition and other domains. LeCun, Yann describes how deep learning has overcome the limitations of the conventional machine learning algorithms and further discusses techniques such as supervised and unsupervised learning with its relation to the multiple hidden layers for a neural network for high level data extraction. Fang, Zhijun proposed a method for abnormal activity detection from video surveillance feed by using multi-scale histogram optical flow on video frames. He improved upon the earlier methods of abnormality detection by applying deep learning network to extract high level features from the video frame. The performance of the proposed work was shown to give effective results when compared to previous methods without deep learning. A more recent application of deep learning for object detection was seen where Smeureanu, Sorina et al used convolutional neural network for detection of abandoned objects in public places to reduce the risk of terrorist attacks which make use of such unidentified objects/luggage He used background subtraction as the first step to obtain static object ie any left bag and then used a cascade of CNN to train his model with various images of luggage obtained from internet. His method was shown to have obtained results better than those obtained with a base CNN model. Larochelle, Hugo, et al discussed the principles for training deep neural networks. They implemented pre-training one layer at a time in a greedy way and then used unsupervised learning at each layer to preserve the information from input. The whole network was fine tuned with respect to the ultimate criterion of interest.
Various methods that can be used for implementation of machine learning on our system were studied. On such widely used model was the tensorflow model. Abadi, Martín et al explained the impact that tensorflow has on deploying machine learning algorithms in the areas of speech recognition, robotics , natural language processing, drug discovery etc. They also described the interface of tensor flow and its implementation on a wide variety of heterogeneous systems such as mobile phones and other high end computational devices such as GPUs. They showed their system to be highly flexible that can be used to express a large range of algorithms for training of deep neural network model. Another model that was developed for implementation of deep learning and neural networks in the Caffe Model by Jia, Yangqing et al . This model was introduced to address the need for computationally efficient and suitable for commercial application of visual recognition. The model was made open source and had bindings with Python and MATLAB. To achieve faster processing, CUDA was used with GPU computation. Many techniques were developed in the past few years as a method for detecting objects in images using deep neural networks. One such algorithm that was studied during the course of this thesis was Single Shot Multibox Detector (SSD) by Liu, Wei et al. SSD discretized the bounding box space into default boxes with different aspect ratios for each feature map obtained. The network then predicted the scores for presence of an object in the default box based on the math with the object shape. SSD was described as a model that is easy to train and resulted in accurate results even for a small input training data of low resolution. The high accuracy in this model was obtained by implementing multiple layers at different scales for prediction. Another literature of importance was the use of deep learning for real time detection and localization of the object. This work presented by Particke, F. et al made use of yet another model for neural network training called You Only Look Once (YOLO). This paper presented an accurate and real time approach for object detection and localization that were used on mobile platforms for optimizing production processes in industry 4.0. The detection of object with YOLO neural network was combined with depth information obtained from RGB-D camera. YOLO was explained as a fast method for prediction of objectness score for each bounding box obtained from the feature map of the image pixels. YOLOv2 divided an image in a grid and predicted bounding boxes for each grid cell with four coordinates and a confidence score for those boxes. The last model studied for the training of neural network was the faster R-CNN model. Ren, Shaoqing et al proposed faster R-CNN model in which they introduced Region Proposal Network (RPN) that used the full-image convolutional network features with detection network. RPN predicted the object bound and prediction score at each position. RPNs were shown as fully-convolutional network (FCN) and they could be trained end-to end specifically for the task for generating detection proposals.
Li, Guanbin et al proposed a method that operated at pixel level instead of patch level in deep convolutional neural network and extracted powerful features for object detection.
Literature survey was also done to study the current trends and evolution of video surveillance technology that was used to procure videos of people smoking in public areas. This research was done to study the available methods for analysis of video in a crowded scene. Liao et al reviewed the current progress made in video-based anomaly detection. They focused on the research work done in finding the feature representation of videos. Work done using deep learning techniques for anomaly detection and action recognition was also reviewed in this paper and explained how deep learning algorithms can help in learning representations from the video data itself. Anomaly detection had been a challenging and active topic of research that involves use of video surveillance to alert authorities on time for any suspicious activity. Javan Roshtkhari et al presented one such approach for detecting suspicious events by using the video itself as the training samples for valid behaviors. These salient events were obtained in real-time in a densely sampled video. They used probabilistic approach to calculate the likelihood of an event being normal in the video and the video frames with very low frequency of occurrence were assumed to be anomalous events. Lu G et al proposed the used of frame-selection framework with unsupervised learning technique for automatic summarization of video content to gain certain perspective of a video stream without having to view the video in its temporal entirety. The proposed technique provided an efficient and reliable solution for the deployment of vision based surveillance systems in public areas that had applications in traffic analysis, crowd monitoring and crime/terrorist activity deterrence.
Feng Y et al presented another approach for automatic representation of video events by extracting motion and appearance features using PCANet. They used deep gaussian Mixture Model (GMM) to model event patterns with observed normal events. Deep GMM stacks multiple layers of GMM which makes it a scalable model. Similar to the paper, this paper also used the probabilistic framework to judge if the detected event should be categorized as normal or an abnormal event.
Mehran, Ramin et al depicted a different approach from the ones discussed above in detecting abnormal activities in a crowded scene. They implemented a social force model in which a grid of particles was placed on top of the image and optical flow was applied on the grid. The interactive forces of the moving particles were estimated using the social force model which was mapped for every pixel of the frame. Spacial and temporal volumes of force flow along with bags of word approach were used to classify the behavior as normal or abnormal. This method was show to be successful in capturing and classifying the dynamic behavior of crowd. Xiao, Tan proposed a technique called sparse semi-non negative matrix at each pixel to learn local patterns. They then constructed a histogram of non negative coefficients (HNC) replacing the previously applied histogram of oriented gradients to detect the local features in an image more expressively. The HNC resulted in a probability model that was used to predict the presence of abnormal events in a video sequence.
A review on the various aspects of video surveillance was presented by A Baumann et al. They provided a systematic review wherein the effectiveness of measures such as segmentation, detection and tracking was compared. Issues such as robustness of the system, normalization etc were considered and a framework was introduced to evaluate the performance of various surveillance systems based on vision. A review paper on the methods of human detection in videos and its applications was also studied to help understand the techniques that can be implemented in object detection in a crowded scene.
Manoranjan Paul et al presented a review on the steps that have been applied by various authors for effective human detection which has applications in abnormality detection, person identification, fall recognition, congestion analysis etc. The techniques covered in this paper were optical flow, background subtraction and filters based on spatio-temporal features. It described that a detected moving objected and be classified as a human either on the basis of texture or on shape based analysis. The paper further discusses methods for background subtraction such as the gaussian mixture model (GMM) and temporal modeling. Since background subtraction was a key component of most of the object detection techniques in video surveillance where videos are captured by static cameras for various applications, a review on the available background subtraction algorithms was studied by Piccardi et al. They reviewed GMM and Kernel density estimation (KDE) models. In GMM, background was demonstrated for every pixel independently as a Gaussian probability density function. The Gaussian appropriation was fitted to n most recent pixel values and a pixel was arranged by determining the probability that the pixel esteem depicts a value distinguishable from the prior pixel values. In KDE model, a function was constructed that gave the probability that a given pixel belongs to the distribution of background pixels. The kernel density estimator distribution was constructed from a sum of kernels. While there exists many techniques for object segmentation from background, smoke detection in videos still remain a complex challenge due to its abstract and dynamic nature. This challenge complicates the task of detection of public smoking in a crowded area using video surveillance.
We provide you with original essay samples, perfect formatting and styling
To export a reference to this article please select a referencing style below:
Sorry, copying is not allowed on our website. If you’d like this or any other sample, we’ll happily email it to you.
Attention! this essay is not unique. You can get 100% plagiarism FREE essay in 30sec
Sorry, we cannot unicalize this essay. You can order Unique paper and our professionals Rewrite it for you
Your essay sample has been sent.
Want us to write one just for you? We can custom edit this essay into an original, 100% plagiarism free essay.Order now
Are you interested in getting a customized paper?Check it out!