Intelligent video analysis for enhanced pedestrian detection using optimized deep convolutional neural network

Abstract

Recent developments in computer vision has enabled the Artificial Intelligence to work on computers to extract, analyse and understand the useful information from an image and video data. Video analysis is an activity of discovering and monitoring the objects, object attributes, object behaviour and its pattern related to the scene captured in the video. Intelligent video analysis has increased its applications across all domains namely home automation, health-care, retail, entertainment, automotive, safety and security. Majority of the applications in video analysis focuses on video surveillance. Pedestrian detection is an essential and significant task in video surveillance. Though there exist many advancements in the area of computer vision, efficient identification of meaningful pattern in the captured scenes for accurately detecting the pedestrian still remains challenging. Pedestrian detection model can be developed using traditional machine learning, hybrid machine learning and Deep Learning (DL) approaches. Pedestrian detection process involves two main steps namely feature extraction and feature classification. A video is a collection of sequential image frames. An image is viewed as collection of features represented as pixels. During feature extraction, the pixels of our target object are extracted from the entire image. The extracted features are then classified either as a pedestrian or non- pedestrian. The traditional filter based approaches for pedestrian detection struggles with number of challenges like presence of noise, illumination and resolutions of the images. The Machine Learning (ML) algorithms are capable of discovering patterns from multi-dimensional and multi variant data without human intervention. Technological developments in the area of ML algorithms helps to detect pedestrians. There are many feature extraction techniques available like Haar wavelets, SIFT, SURF, HOG, etc. Each of these feature extraction technique has its own advantages and disadvantages based on the data it is handling. These exists two broad category of ML algorithms namely supervised and unsupervised algorithms. Since pedestrian detection is a classification problem it is necessary to depend on the supervised ML algorithms for classifying pedestrians. Among the different supervised ML algorithms, Naïve Bayes (NB) and Support Vector Machine (SVM) perform well in the process of classifying the input features as pedestrians or non- pedestrians. These ML algorithms manages to perform well on small dataset with simple features. These ML algorithms struggles to perform well on a large image dataset with complex and varying features. The objective of this thesis is to enhance the accuracy of the pedestrian detection from the input video using ML and DL approaches. Initially the pedestrian detection is carried out traditionally using histogram of gradients for feature extraction and SVM for classifying the features from the input video. The benchmarking datasets from VISOR and INRIA image repositories were used for model training and model validation. VISOR dataset is made up of simple images with uniform features whereas image frames extracted from INRIA dataset is made up of complex images with varying features. newline