CNN Pre-Trained Model Performance Analysis For Vehicle Detection On Low Illumination

In this study, we compared and analyzed the performance of CNN pre-trained model for vehicle detection in badly illuminated situations and using moving cameras. The comparison was performed to analyze the ability of the benchmarking CNN pre-trained architecture to perform under low quality images. We conducted experiments using five CNN architectures, namely Faster R-CNN, SSD-ResNet, YOLOv3, YOLOv4, and EfficientDet. The dataset used is a dataset that is taken independently by using a camera mounted on the observer's vehicle (data taker). The results of our experiment show that YOLOv4 used with a mean average precision (mAP) of 0.37 has the best performance compared to the other architectures. While the lowest performance is shown by Faster R-CNN and SSD-ResNet with mAP value of 0. YOLOv4 has good results because its architecture includes steps to recognize objects in various degradation conditions.