Methods and algorithms for detecting and classifying moving objects based on the evaluation of an object’s geometrical parameters in an image sequence for smart lighting systems

Abstract

Modern machine learning based object detection methods have high accuracy; however, at the same time, they require the increased computing power of processing units. Thus, widespread low-performance IoT devices generating a substantial amount of data cannot apply corresponding machine learning algorithms for real-time data processing due to a lack of local computational resources. The Dimensional Based Object Detection (DBOD) algorithm for low-performance single-board computers is developed and evaluated in the course of this dissertation. The proposed algorithm exploits geometrical features of objects in an image and real-world scene parameters (e.g. camera’s focal length, height and angle of installation) as classification features. Extraction and classification of these features are computationally simple procedures that single-board computers can execute in real-time. The algorithm is focused on detecting and classifying objects that are the most expected in urban environments: pedestrians, bicyclists, and vehicles. The algorithm can be applied for processing video sequences captured by a CCTV camera. A method for fast generating synthetic training features for the DBOD has been proposed. The algorithm DBOD has been tested on real and synthetically generated datasets. The results have shown that low-performance systems, such as popular Raspberry Pi, are capable of object classification with the required frame rate and accuracy for smart city applications.