Enhanced Object Detection for Autonomous Vehicles Using Modified Faster R-CNN with Attention and Multi-Scale Feature Fusion

Abstract

The progress of self-driving technology necessitates more stringent demands on object detection systems, and traditional methods are difficult to meet real-time and high-precision requirements in dynamic scenes. Therefore, this study proposes an improved Faster R-CNN model tailored for vehicle object detection in autonomous driving scenarios. Specifically, an enhanced Convolutional Block Attention Module (CBAM) is integrated into the backbone network to strengthen feature representation. The Region of Interest Align (ROI-Align) is employed to improve localization accuracy, especially for small or occluded targets. Moreover, Soft Non-Maximum Suppression (Soft-NMS) is adopted to reduce false negatives in dense object scenarios. Additionally, a multi-scale feature fusion mechanism is introduced to enhance detection performance across varied object sizes. The experiment outcomes indicate that the detection accuracy of the improved model reaches 98.13%, with a miss rate of less than 1.00%. In dense target scenes, the retained accuracy is 94.16%, and the standardized mean square error of target localization is 0.014.In complex environments, the average accuracy of the model in lighting changes, severe weather, and dynamic interference scenarios is 80.45%, 77.83%, and 75.11%, respectively, which is superior to the comparison methods and demonstrates higher robustness. This study enhances the detection performance of faster region-based convolution neural network in automatic driving through technical modifications, solves the problem of feature extraction and target location in complex scenes, and provides important support for the perception reliability of auto drive system.