With the continuous improvement of urban greening, it is more complex and heavy to clean the fallen leaves. According to the characteristics of variable shape, complex background and uneven distribution of fallen leaves, we propose a fallen leaf detection algorithm, which integrates an attention context(AC) network and YOLOv3, named AC-YOLO, to detect fallen leaves on the road quickly and accurately. To solve the problem that small leaves are difficult to detect, we propose the AC network. We use different feature levels as the context information of small leaves and attention mechanism to suppress the influence of complex background and bottom noise, so as to improve the ability of small leaves detection. In addition, we use the Mish activation function to replace Leaky ReLU to enhance the generalization ability of the model and improve the accuracy of leaf detection. Finally, considering that the stacking of fallen leaves has an impact on the work efficiency of the cleaning robot, we propose the non-maximum fusion to fuse the dense fallen leaves detection boxes, which promotes the detection of dense fallen leaves and improves the work efficiency by reducing goal nodes. Experiment results show that the cover of leaf detection based on the AC-YOLO algorithm is 95% and the detection speed is 53 frames per second, which can complete the leaf detection task in the practical application and realize the efficient and intelligent leaves cleaning.