YOLOv5s FMG: An Improved Small Target Detection Algorithm Based on YOLOv5 in Low Visibility

Abstract

Accurate and real-time detection of small targets of pedestrians and cars in video images is indeed crucial for various applications such as autonomous driving and urban management. Existing detection algorithms face challenges related to small targets and low visibility, resulting in issues such as low accuracy, missed detection and reduced detection efficiency. This paper proposes an improved YOLOv5s FMG (Fine-tuning Slice, Multi-spectral Channel Attention, Ghost Bottleneck) detection method based on YOLOv5, which firstly introduces fine-tuning slicing aided hyper inference (SAHI) to generate small target objects by slicing the pictures into the network. Secondly, the multi-spectral channel attention (MCA) module is integrated into the feature extraction network, which enhances the information dissemination among features and strengthens the network's ability to distinguish between foreground and background. Then, the network uses the convolution network to extract features instead of the full connection layer and uses the lightweight Ghost Bottleneck instead of the bottleneck structure. Finally, the prediction part adopts the complete intersection over union (CIoU) loss function to achieve accurate bounding box regression. Based on the experimental results conducted on the self-made dataset, compared to YOLOv5s, the mAP (0.5) of YOLOv5s FMG on the dataset is improved by 9.3%, and the mAP (0.5:0.95) is improved by 2%. At the same time, the frames per second (FPS) is increased by 41.8%, and the number of parameters has been reduced by 18.5%. The proposed method demonstrates successful detection of small targets of pedestrians and vehicles, ensuring its effective applicability under conditions of low visibility.