Abstract The ability of vehicles to detect obstacles on the road is a critical component in advancing autonomous driving systems. Driving involves complex perception and decision-making, which is a challenge for humans and automated systems. In this work, we present and evaluate two object detection models for identifying various road entities, including cars, trucks, pedestrians, and other obstacles. The first model is a modified Region-based Convolutional Neural Network (R-CNN), and the second is a single-stage detector based on the EfficientDet-D0 architecture. In this work, R-CNN uses VGG-16 as its base CNN model for feature extraction. The combination provides strong representational power by utilizing VGG-16 with structured region-based detection from R-CNN, enabling accurate obstacle classification and localization. The R-CNN model was enhanced with architectural modifications tailored for two-stage detection using hybrid fully connected layers (FCL). In contrast, the EfficientDet-D0 model was trained using transfer learning on the Udacity self-driving car dataset. The EfficientDet-D0 model demonstrated superior performance in real-time conditions, reporting a detection accuracy of 76.8% mAP@0.5, an IoU of 0.73, and a processing speed of 30 frames per second (FPS). In contrast, the custom R-CNN model achieved 69.3% mAP@0.5, with a notable processing rate of 32 FPS, making it suitable for real-time deployment. Despite the promising results, certain obstacle categories remain inadequately detected at high vehicle speeds. Our model detects road obstacles in real-time, achieving both low latency and high accuracy. We present the accuracy and loss metrics for both models to provide a detailed analysis of their performance compared to baseline methods. The ablation study demonstrated that transfer learning significantly enhanced model performance. The EfficientDet-D0 model’s detection accuracy dropped by over 14% without transfer learning.
Veningston et al. (Thu,) studied this question.