Safe and Efficient Traffic Obstacle Detection Using Fine-Tuned Modified Faster R-CNN

The proposed Fine-tuned Modified Faster R-CNN (Region-Based Convolutional Neural Network) implementation system uses computer vision and deep learning to improve the effectiveness and safety of traffic obstacle identification. This automated system uses machine learning algorithms based on the ResNet-50 architecture. A dataset of images of bikes, pedestrians, traffic signals, and cars, are used for training and testing. Accurate traffic element identification and localization are made possible by the Faster R-CNN object detection model. The model's performance is improved by the application of strategies including Stochastic Gradient Descent, Cross-Entropy, and Fine-tuning, yielding an accuracy rate of 98.26%. In real-time circumstances, the model creates precise bounding boxes around things it detects, opening the door to a more automated and secure transportation system. The proposed model is compared with other object detection models like Fast R-CNN, R-CNN, YOLOv5, SSD, RetinaNet based on five parameters: Accuracy (mAP), Latency, Model Size, Dataset, Bounded box loss. The comparative analysis shows superior accuracy compared to existing systems.