Rotating Frame Labeling and Improved YOLOv7-based Algorithm for Vehicle Object Detection

This paper proposes a novel rotating frame labeling method based on the bottom of the vehicle to overcome the limitations of horizontal labeling frames, which often contain unnecessary background information and overlook vehicle directionality. The proposed method utilizes the CARLA simulation simulator for dataset construction. To enhance compatibility with the dataset, we introduce an improved rotating vehicle target detection algorithm based on YOLOv7. The algorithm incorporates an adaptive feature fusion module to effectively utilize features at different scales, giving priority to features with significant information through adaptive weight learning. Additionally, the integration of CBAM (convolutional block attention model) allows the model to fully exploit channel and spatial features. Furthermore, the Transformer module is integrated to capture global and contextual information, thereby improving the network's detection capability. To address the variability in detection frame angles within the dataset, we replace the original loss function with KLD(Kullback-Leibler Divergence) loss for target rotation frame regression. Experimental results demonstrate excellent performance on our self-built dataset, achieving a mAP value of 94.62% and an FPS value of 37. Moreover, the combined results outperform other algorithms on the HRSC2016 rotating target detection dataset.