Pedestrian Detection Surveillance Videos using Background Modeling and Feature Engineering

Considering the humongous amount of video data being produced everyday by countless number of surveillance cameras,it has become imperative for the scientific community to automate the video analysis and let the machine detect and track the objects of interest. Most of the surveillance applications revolve around human activity monitoring, where detection of pedestrian (or human or people or person) serves as the precursor for the subsequent analysis. This dissertation focuses on detecting pedestrians in video data for a surveillance framework across a stationary camera view. The research includes motion information as a prior cue to extract all moving objects followed by the design of a feature descriptor to classify them as either pedestrians or non-pedestrians. A comprehensive background model is presented to detect the set of moving objects across a static camera view. A hybrid color space is suggested for suitable pixel representation. The sample variation of pixel sequence along the temporal domain is taken into consideration to model a multi-modal background. An outlier labeling methodology is employed to identify the potential foregrounds, whereas the frequency of objects’ appearance is applied to update the modeled background. A feature extraction methodology is presented to yield a generic signature to represent pedestrians. A holistic shape descriptor is proposed to encode the human body using weighted orientation histogram. The Golden ratio of human anatomy is employed to develop apart-based detector to address the problem with partial occlusion and unusual articulation. The choice of an appropriate classifier for pedestrian classification is decided on the basis of its behavior in terms of the rate of change in accuracy with varying the model parameter. This research also includes the design of a fully convolutional deep architecture to detect pedestrians by automatically selecting the desired region proposals as well as learning the requisite feature representation. The architecture facilitates end-to-end training and thereby improves the overall performance eliminating the bottleneck caused by the multi-stage pipeline structure of conventional machine learning. A two-stage approach is suggested to separate the human-look-alike hard negative backgrounds from the actual pedestrian objects. Besides, feature maps from multiple intermediate layers of the underlying network are taken into account for small-scale detection. Exhaustive simulation has been carried out to analyze each of our propositions both separately and as a unified framework. A set of standard evaluation metrics has been taken into consideration to compare the proposed models with their counterparts