Real-time processing and analysis of truck blind spot images based on computer vision

Due to factors such as the large body volume, high position of the driver's cabin, and limited range of rearview mirror reflection, trucks have blind spots when making right turns and reversing. This prevents drivers from fully assessing the surrounding environment, leading to safety incidents. This paper proposes a computer vision-based method for detecting obstacles in truck blind spots. Deep learning-based detection techniques, particularly the YOLO series algorithms, are commonly used due to their simplicity and speed in real-time detection. However, inaccuracies in output bounding boxes, especially in detecting small-scale and overlapping objects, lead to low detection accuracy. This paper improves upontheYOLOv5s framework by incorporating attention mechanisms to enhance the feature extraction capability for small-scale and overlapping objects. Additionally, to prevent image feature loss and resolution reduction caused by pooling operations, multiple parallel dilated convolutions with different dilation rates are used instead of pooling layers, forming an atrous pyramid pooling structure. This structure is added after different hierarchical feature outputs and fused to increase the network's receptive field.