Research On Lightweight Object Detection Algorithm For Vision Guided AGV

With the rapid development of "Made in China 2035",intelligent warehousing and logistics systems,freight terminals,and even daily life scenarios,Automated Guided Vehicle(AGV)play a crucial role.With the continuous development and application of computer vision technology,visual guided AGV object detection methods have become a research hotspot,with a focus on improving model accuracy while reducing the number of model parameters to ensure the smooth deployment and operation of the model on the mobile end.The main research content is based on deep learning models,such as YOLOV4 and YOLOV5,from the perspectives of object recognition and security,lightweight object detection algorithms are designed for daily life and traffic scenarios,respectively.Subsequently,the improved algorithms are pruned to facilitate deployment on visual guided AGV.The main work of this article is as follows:(1)The object detection algorithm designed for material handling AGV and disinfection temperature measurement AGV,which are commonly used in daily life scenarios,is from the perspective of object recognition to facilitate the safe driving of visually guided AGV in airports,docks,and hospitals.Considering the high complexity of the YOLOV4 algorithm model and its difficulty in deployment on mobile devices,a lightweight object detection model based on cross channel fusion and attention mechanism is proposed.The algorithm is based on YOLOV4.Firstly,the MV1 module and CDC module are reconstructed,and a lightweight model based on deep separable convolution(DSC)is designed;Secondly,while expanding the network neck,a cross channel path is fused to enhance feature extraction;Finally,an efficient channel attention mechanism is introduced to enhance the detection ability of the model.The experimental result show that compared with the newer YOLOV7,the improved algorithm reduces the number of model parameters by 60% and improves accuracy by 0.74% on the Pascal VOC07+12 dataset,fully demonstrating the effectiveness of the algorithm.(2)In response to the problem of overlapping multi object information collected by AGV in complex road backgrounds,which can lead to false positives and missed detections,we propose a object detection algorithm that integrates the Ghost module and attention mechanism from a security perspective to avoid collisions between AGV and people and vehicles during movement.This algorithm is based on the YOLOV5 n network and redesigns the C3 modules for the head and neck;At the same time,SE(Squeeze and Exception)attention mechanism is added in front of the prediction end to make the algorithm focus on the effective position of the object;Finally,removing Mixup data for enhancement during the training process is more suitable for training and detection of small models.The experimental result show that the improved algorithm has a model accuracy of 93.28% on the KITTI dataset,which reduces the parameter count by18% compared to YOLOV5 n.The model only occupies 6.13 M of memory,creating conditions for mobile deployment.(3)In order to further solve the problem of difficult deployment of object detection algorithms on mobile devices and excessive reliance on hardware conditions,we propose a channel pruning strategy for lightweight object detection models.By constraining the regularization coefficient of the BN(Batch Normalization)layer of the convolutional network,the parameters of each BN layer approach 0,so as to cut off the unimportant channels.By designing comparative experiments with different pruning rates to select parameters,the channel pruning of the improved algorithms in Chapter 3 and Chapter 4 was completed.The model parameter count was reduced by 26% and 21%,respectively.After fine-tuning the parameters,the recovery accuracy was retrained,demonstrating the effectiveness of the pruning strategy.