The Affect of Image Lighting in Determining Anchor Box for Vehicle Object Detection using Faster R-CNN

Department of Informatics, Sanata Dharma University, Sleman, Yogyakarta, Indonesia

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Effective traffic management is a critical component in urban safety and the efficiency of road use. Modern traffic system to manage traffic uses the ability to detect traffic to better adjust traffic flow. Object detection can be used in such cases, where CCTV feed is used as a reference to locate and count vehicles. Data collection was carried out during both day and night situations. Faster R-CNN is one of such algorithms that has proven to be robust for object detection. It uses a two-stage process that makes use of anchor boxes to determine the existence of objects in an image. This research aims to find out the effects of differing lighting conditions on the road on the effect of the anchor box used for the model to get the best result. Under both day and night conditions, the best models used anchor size of [64x64; 128x128; 256x256; 512x512] and anchor ratio of [0.5; 1.0; 2.0]. Lighting conditions does not seem to affect the choice of the anchor box set. It is found that smaller and more varied anchor sizes lead to lower RPN loss. While a more diverse set of anchor ratios provides smaller detector loss.