Research On Detection Method Of Complex Fallen Leaves On Road Based On YOLO

With the continuous improvement of urban greening,fallen leaves cleaning work has become more complex and heavy.The fallen leaves have seasonal characteristics and need to be cleaned repeatedly.Especially in the park environment with high greening rate,they can only be cleaned manually,which is time-consuming,laborious and inefficient.In recent years,with the rapid development of artificial intelligence and target detection technology,it is a new way to study the intelligent detection method of fallen leaves and realize the intelligent cleaning of fallen leaves in the field of sanitation.Aiming at the problems of complex background of fallen leaves on the road,many small and medium-sized leaves and uneven distribution,this thesis proposes ACNet and Non-Maximum Fusion algorithm to improve the existing methods,and carries out model pruning to realize model lightweight.Details are as follows:1)In order to achieve the accurate detection of fallen leaves on the road,this thesis firstly establishes a fallen leaf dataset based on the characteristics of leaves.Then,this thesis proposes ACNet to solve the problem that small and medium-sized fallen leaves are always missed.The ACNet fuses different levels of feature mappings as the contextual information of small targets,and introduces a self-attention mechanism to suppress the effects brought by complex backgrounds and underlying noise,so as to improve the detection capability of small fallen leaves.Finally,the Mish activation function is used to replace Leaky Re LU to improve the generalization ability of the model and improve the accuracy of leaves detection.Experiments show that the improved algorithm is more efficient on fallen leaves detection than the original algorithm.2)To solve the problem of difficult detection of dense fallen leaves and high miss rate,as well as to improve the navigation efficiency of the cleanup robot,this thesis proposes a back-end processing algorithm,Non-Maximum Fusion(NMF).NMF predefines δ to scale the high confidence prediction boxes to fuse the intersecting or adjacent boxes instead of suppressing them.Experiments on the fallen leaves dataset show that NMF significantly improves leaves detection coverage with a coverage rate of 95.3%.In addition,NMF greatly reduces the number of target nodes for path planning.Since NMF works at the back end of the detector and does not require retraining,it can be simply integrated into the leaf detection algorithm.3)In order to apply the falling leaf detection algorithm to edge-devices,structured pruning of AC-YOLO is performed to achieve model lightweighting.In this thesis,we use the scaling factor of the Batch-Normalization layers to determine the importance of the channels to implement model pruning.We set selective pruning to skip the Shortcut directly connected layer to solve the channel mismatch problem.In addition,a local threshold is set to ensure the minimum number of channels in the convolutional layer,which guarantees the connectivity of the network.The final experiments show that Light-AC-YOLO achieves a detection speed of 80 frames per second and a leaf coverage rate of 87.2% on the falling leaves dataset,which has better real-time performance with guaranteed accuracy.To sum up,the algorithm proposed in this thesis has a good effect on fallen leaves detection and can solve the problem of complex falling leaves detection on the road.This thesis has 42 figures,13 tables and 90 references.