Real-Time Deep Learning-Based Object Detection Framework

Recently real-time detection, and recognition of an object of interest are becoming vital tasks in visual data processing and computer vision. Various models have been deployed to implement object detection and tracking in multiple fields. However, conventional classifiers are often faced with challenging tasks that visual frames come distorted due to overlapping, camera motion blur, changing subject appearances, and environmental variations. Models using OpenCV-based HAAR feature-based cascade classifiers, without integrating any error minimizing object detection algorithm, were unable to accurately detect an object and track it in a changing environment. Therefore, developing an embedded powerful framework for realtime object detection and recognition becomes more of a vital need for future implementation in various fields. This study presents a powerful technique for a real-time detector that utilizes integrated Deep Learning Neural Networks (DNN) for optimal computational accuracy. Deploying such a framework will ensure the flexibility and reliability of the detector by eliminating the sources of distortion previously mentioned. The model relies on integrating the ImageAI deep learning libraries and You Only Look Once (YOLO-v3) object detection method with a DarkNet53 architecture. The algorithm was trained using the TensorFlow framework to ensure accurate data processing. This paper targets one vital component of our long-term project of developing a multi-agent system, as the proposed model is to be implemented in autonomous agents for the detection of landmines, ocean debris, and wildlife beside environmental scanning missions. In this study the performance of the model has been assessed through detecting and collecting tennis balls as a preliminary test for real-world applications. The model was able to approach the desirable result of surpassing the accuracy of conventional detectors.