Research On Target Detection Optimization Algorithm Based On YOLO

With the progress of science and the continuous development of deep learning,Target detection technology is becoming more and more mature.Pedestrian detection is one of the important areas of Target detection technology.This thesis mainly studies the Target Detection Optimization Algorithm Based on YOLO(You Only Look Once).Aiming at the problem of poor detection effect under the condition of insufficient light in the daytime and no light at night,the YOLO algorithm is optimized respectively.The specific contents are as follows:To solve the problem that YOLO cannot accurately identify pedestrians in the daytime with insufficient light,the Bimodal Pedestrian Detection Algorithm Based on YOLO is improved in this thesis.In order to enrich the characteristics of the target to be detected,after the modal characteristics of visible light and infrared light are extracted from the trunk feature extraction network,the extracted features are weighted and fused by the dual-mode feature weighting module.Then,the integrated features are optimized through the enhancement of channel attention mechanism and spatial attention mechanism to improve the representativeness of important features.Experimental results show that even in the absence of illumination,target occlusion and multi-scale environment in the daytime,the proposed algorithm still has a good detection effect and improves the detection accuracy with an average accuracy of 84%,and the detection speed is basically the same as the original algorithm.Aiming at the problem that visible light images cannot be used to detect pedestrians in the dark environment at night,the Infrared Pedestrian Detection Algorithm Based on YOLO is improved in this thesis.This thesis improves the trunk feature extraction network Darknet-53 in two aspects.On the one hand,the deconvolution layer is added into the residual module to transform small-scale features into large-scale features for extraction,so as to improve the detection effect of small-scale targets.On the other hand,the connection mode of Darknet-53 was improved to promote the feature extraction capability of the network.In this thesis,Io U is also improved to enhance the sensitivity of position deviation between the prediction box and the real box in detection,and to reduce the training time of the network by optimizing activation function.Experimental results show that the proposed algorithm not only improves the detection accuracy,the average accuracy reaches 93%,and can effectively detect small scale pedestrian targets in the dark environment at night.