Development of object detection technique using machine learning

Abstract

In the era of automation and advances in technology, the area of Image Processing and Computer Vision has grown rapidly. Different Object Detection applications include face detection, pedestrian detection, human behavior analysis, image retrieval, photo tagging of faces, autonomous driving, and video surveillance. Object Detection describes the problem of locating different objects (object localization) and assigning category (object classification) to each object present in images or videos. Machine learning (ML) is a field of Artificial Intelligence (AI) that deals with the study and design of algorithms and models that computer system uses to perform a certain task. The traditional ML techniques for image classification are made up of two sub processes namely the feature extraction process and classification process. In literature Scale Invariant Feature Transform (SIFT), Haar-like features, Histogram of Oriented Gradients (HOG) are utilized as feature extractors and Support Vector Machines (SVM), Deformable part models (DPM), Adaboost as the classifiers. Illumination conditions, different poses, occlusions, small datasets, class imbalance are different challenges that make object detection difficult. Deep learning techniques have automated the feature extraction process further leading to improved performance. Object detection algorithms have gone through significant improvement since 2012, due to the availability of datasets, Graphical Processing Unit (GPU) and pretrained networks. Neural networks with many layers have a higher capacity to learn complex features and can optimize certain tasks. The research using convolutional neural network(CNN) is still evolving but still there are challenges to be handled such as occlusions, variations in pose and viewpoint, object rotation and scale changes, different illumination conditions, different object sizes, distortion, intraclass variation which are faced by object detection algorithms in different applications. The main problem while detecting the traffic lights