ROI And Multi-block Local Binary Patterns For Robust Pedestrian Detection

Over the past20years, pedestrian detection has been one of the popular research top-ics in computer vision and machine learning owing to its potential usages such as driving assistance, traffic control, security supervising, and pedestrian circulation analysis. Present smart vehicles require a protected function which is capable to detect and track the pedes-trians thereby preventing or limiting the occurrence of accidents. Even though there are currently plenty of pedestrian detection systems designed by many research institutes, not many of them can handle the motion and real time detection challenges. Additionally to these two problems, there are some common challenges such as:occlusion, cluttered image background, illumination variation, and pedestrian postures.The objective of this work is to develop a rapid and robust pedestrian detection model, machines learns patterns from public datasets and categorize them into the desired classes in a basic and efficient manner:(pedestrians and non-pedestrians decisions). This dissertation formulates four major contributions:firstly, to detect moving individuals in the videos, a meaningful "region of interest (ROI)" method is proposed. This method lowers the compu-tational cost and boosts the detection speed. Present researches show that sliding window classifiers are the widely used techniques in pedestrian detection due to their superior perfor-mance. Sliding window considers all possible sub-windows of the image and the classifier makes a final decision whether or not they detect a pedestrian. Searching potential candidate from whole image is computationally expensive given that the size of the scene gets larger. The searching technique is inadequate in real time pedestrian detection, the searching area must be reduced. However the "region of interest" phase will significantly reduce the number of detection windows needed to be considered. Secondly, multi-block local binary patterns (MB-LBP) were built as descriptor for human detection. The MB-LBP descrip-tors are computed to encode the neighbouring appearance and details of human objects, the obtained feature patterns at fixed resolution are collected into a high dimensional feature vector. They are created to be strong with small variations in image illumination, colour, contour locations and directions while staying remarkably discriminative for overall visual form. Neighbouring regions are needed for good performance. To enable the human de- scriptor to become stronger, a new textural feature, namely "spare MB-LBP" is developed. Thirdly, a cascade classifier was constructed, it handles significant information. Despite the fact that, the Adaptive Boosting (Adaboost), Neural Network, and support vector machine (SVM) classifiers present nice quality detection results, they are not yet able to deal with real time systems, leading to long processing time. In this dissertation, consideration is given to the cascade-improved fisher linear discriminant analysis (IFLDA) for classification, which incorporates accurate and fast computation speed benefits (boost the speed of the detection system). Finally, an enhanced part based method is presented for person detection, which first detects local body-parts such as heads, torso, legs, and then fuses them to generate an entire human detector.To test the feature sets, a comparatively easy learning framework is used to classify every possible image region as a pedestrian or as a non-pedestrians. This work focuses on pedestrian detection in still images and videos, it can be applied for smart cars and video surveillance. Experimental results working with a number of standard datasets prove that the proposed approach performs favourably in comparison to some of the state-of-art methods.