An Optimization to YOLOv3-tiny For Real-Time Detection of Small, Fast Moving Objects

International Journal of Research and Scientific Innovation (IJRSI) | Volume VI, Issue XII, December 2019 | ISSN 2321–2705

An Optimization to YOLOv3-tiny For Real-Time Detection of Small, Fast Moving Objects

Pavan Shiralagi¹, Rahul Bhandary², Rajeshwari B³, Bajarangbali R⁴

^1,2,3,4Department of ECE, PES University, Bangalore, India

Abstract— A project to obtain tennis statistics based on tennis ball tracking led us on a search to find real time object detection on small, fast moving objects. Realizing there were no methods available that satisfied our requirements, we optimized the detection method that came closest, YOLOv3 tiny [You Only Look Once Version 3 Tiny], to come up with YOLOOv3 tiny [You Only Look Once, Optimized]. YOLOv3 tiny is the third iteration of a Computer Vision application that is used to detect objects in real time. However, it is limited by the size and speed of the object relative to the camera’s position along with the detection of False Positives due to incorrect localization. In this paper, we explore optimization techniques to extend the use of YOLOv3 tiny to accurately detect small, fast-moving objects. The techniques discussed in the paper were tested on detecting a tennis ball moving up to 160kmph with a minimum angular size of 0.414 degrees at a rate of 30 frames per second. The accuracy was found to be 95.268% on a video containing 4600 frames sampled at 30 frames per second (assuming the object is always in the frame) on a GeForce 1050 Graphics Processing Unit. This optimization includes a method of elimination of false positives to increase accuracy.

Keywords— Real Time; Small Object; Fast Moving; False Positives; Optimization

I. INTRODUCTION

Computer Vision has been one of the fastest developing fields in the recent past, with new methods for faster and more accurate object detection leading the growth. However, there is still much room for the improvement of detections on small and fast-moving objects relative to a camera. Many methods such as R-CNN [1], fast R-CNN [2], faster R-CNN [3], SSD [4] and YOLO [5] have been proposed for object detection. However, these methods struggle with the detection of small objects [6], primarily because of the smaller resolutions of the final layers of the neural network, which cause the features of the small object extracted at lower layers to become negligible at the final layers, leading to the problem of false positives where objects that may bare a slight resemblance to the object of interest are detected as the object of interest.

Read More