Publication

Abstract : Small-size object detection (SOD) is one of the challenging problems in computer vision applications. SOD is highly useful in defense, military, surveillance, medical, industrial and analysis in sports applications. Various algorithms were developed in the past to solve the problem of SOD. However, the algorithms developed are not suitable for real-time applications. In this work, a convolutional neural network architecture based on YOLO is proposed to enhance small objects' detection performance. The proposed network is inspired by the ideas of Residual blocks, Densenet, Feature Pyramidal Network, Cross stage partial connections, and 1 × 1 convolutions. The Receptive field and the reuse of feature maps are the main factors in the design of the architecture and is hence referred to as RFSOD. It is developed as a lightweight network to suit real-time applications and can run smoothly on single-board computers such as Jetson Nano, Tx2, Raspberry Pi and the like. The proposed model is evaluated on various public datasets such as VHR10, BCCD dataset and few small-size objects from the MS COCO dataset. This work is motivated by the need to develop a vision system for a badminton-playing robot. Therefore, the proposed model is also tested on a custom-made shuttlecock dataset. The model's performance is compared with the state-of-the-art deep learning models that are suitable for real-time applications. The hardware implementation of the proposed model was carried out on Jetson Nano, Raspberry Pi4 and a Laptop with an i5 processor. Improved Detection accuracy was observed on small objects. More than 2 × detection speed was obtained on Raspberry Pi, and i5 processor while 30% improvement was observed on Jetson Nano with real-time videos.