Automated Identification Of Pedestrian' Attributes For Behavior Analysis In Surveillance Systems By Employing Deep Learning Techniques

Video surveillance is emerging as one of the demanding technology in maintaining security nowadays. As this technology requires the continuous monitoring by watch persons, which becomes a boring job and many times the human watcher misses some important events which can lead to many security issues. For this reason,video surveillance technology is moving towards video analytics in which the video streams are automatically analyzed by computer algorithms. Analyzing people behavior using video analytics is one of the hot topics nowadays. The study has been conducted to investigate better solutions for Automated Identification of Pedestrians' Attributes for Behavior Analysis in Surveillance Systems by Employing Deep Learning Techniques. In this thesis, we propose methodologies for five important pedestrian attributes that help in analyzing human behaviors in automated environments. These attributes involve (a) pedestrians' full body orientation(In which direction, a pedestrian is moving), (b) pedestrians' direction of attention/Head pose estimation, (in which direction, pedestrian is watching), (c)distance and dimensions (height and width) of pedestrians from camera field of view(FOV),(d) identification of pedestrian,himself, as a "pedestrian attribute" and (e)Pedestrians gender recognition (Either Male or Female). The summary of the methodologies used to address these attributes is as followsPredicting Pedestrians' direction of movement and intentions has been an entreating application in human action acknowledgment situations. While moving in one course,people on foot may have their visual considerations in different bearings.The investigation of such problem by means of computer vision is at some point alluring for pedestrian behavior study. This work focuses on pedestrians' head-pose and full-body direction estimation by using a deep learning approach. A convolutional neural network model is exhibited as deep learning methodology for orientation prediction. Two different datasets are set up for the two types of orientation calculation.We prepare ours model independently on the two arranged datasets having eight orientation views. We perform testing on existing data sets and on video sequences taken by us in various scenarios. The execution comes about show that our system successfully characterizes head poses and body views various environments.Pedestrian classification via computer vision has been the theme of enthusiasm over past numerous years. Different conventional and neural system inspired methodologies have been utilized to perceive people on foot. we utilize stacked sparse autoencoder for extracting the features that contain pedestrians' objects. Salient feature maps of the images with pedestrians are produced with the assistance of SLIC superpixel generation by using graph manifold ranking method. The salient feature-maps are afterwards, supplied to the stacked sparse autoencoder. The end classification is performed by passing reconstructed data (by stacked autoencoder) to Softmax classifier.The distance and dimensions calculations of walking people through cameras in real-time environments are required in many situations It is alluring to have a non-contact estimation structure. We propose a framework that uses simple mathematics to calculate the distance and dimensions of pedestrian objects that are within the field of view of a monocular camera. Before estimations, single-shot learning is employed to adjust camera per the environment. For single shot-learning an L-shape marker is utilized by putting it first at least distance and after that at a generally far distance from the camera. The corner points of the marker are calculated at the two placements and through simple line equations, per pixel length at object placement is estimated. The background subtraction strategy through mean filter is used to get foreground moving object. These foreground objects are then predicted as pedestrian and non-pedestrian with the help of convolutional neural network based classifier. Finally, the distance and dimension of the pedestrian object estimated with the help of reading taken at single shot learning step.Pedestrians' gender is a soft attribute which is useful in many areas of computer vision. Apart from its importance, the researchers think pedestrians' gender prediction as one of the toughest methodology in computer vision. In this work, we propose an approach that utilizes a deep learning based approach to classify a pedestrian as a male or a female. As a pre-processing step, we perform pedestrian parsing by existing deep decompositional neural network approach. The output of the network is the binary mask that maps the pedestrian full body from the input image. The pedestrian body image is extracted by applying the generated pedestrian mask to the input image. This pre-processed image is then passed to the proposed stacked autoencoder with softmax classifier for classification.We present another deep learning based approach to analyze the pedestrian gender. we utilize deep convolutional neural network as a deep learning tool for gender predictions. The proposed approach first extract the parsed image of the pedestrian(produced with the assistance of existing deep decompositional neural system based approach). The parsed pictures with extracted foreground pedestrian object are then isolated to full body and abdominal area images. These two sorts of images are then provided independently to the proposed convolutional neural system for gender forecasts. The full body based predictions are grouped under frontal, back-perspectives and mixed pedestrian's views. The abdominal area pictures are arranged under eight classifications with respect to abdominal area apparel.The performance measures and comparisons with existing works depict the robustness and applicability of our proposed methodologies.