FOS-YOLO: Multiscale Context Aggregation With Attention-Driven Modulation for Efficient Target Detection in Complex Environments

Computer vision techniques have significantly advanced autonomous driving but still face specific challenges in complex environments, such as accurately detecting small, occluded, or low-contrast objects amidst dynamic backgrounds and varying lighting conditions. This article introduces FOS-you only look once (YOLO), a novel model designed to enhance detection performance in these scenarios. FOS-YOLO integrates the FocalNet mechanism to improve feature representation, effectively addressing target scale variations and low contrast through multiscale context aggregation and multilevel modulation. In addition, the zoomed spatial convolutional block attention module (ZS_CBAM) is introduced to improve the detection of small targets, occlusions, and illumination changes by effectively fusing channel and spatial attention with a focus on scaled spatial features. The model also includes two lightweight modules, the adaptive depthwise enhanced module (ADEM) and the lite enhanced reduction module (LERM), which reduce parameters and computational load, accelerating convergence and improving accuracy. Experimental results, compared with state-of-the-art (SOTA) methods, show that FOS-YOLO achieves a mAP of 90.4% on the KITTI dataset and 64.3% on the RTTS dataset, with a 20.07% reduction in parameters, significantly enhancing real-time detection accuracy and efficiency.