Research On The Detection Of Multi-type Targets On Road Surfaces Based On Deep Learning

Target detection is to detect the target of interest from the input information.Whether the target detection algorithm can accurately and timely detect the target is an important indicator to judge the quality of the target detection algorithm.With the continuous development of deep learning technology,target detection algorithms based on deep convolutional neural networks have been widely used in many fields such as artificial intelligence and information technology.In this paper,the target detection algorithm based on deep learning is applied to the detection of multiple types of targets on the road surface,and the research is mainly carried out from the following aspects:(1)Aiming at the problem of incomplete data sets in the current multi-type target detection field,this article analyzes the public target detection data set,combines the TT100 K data set and the collected image data,and uses data annotation software to analyze the target information in the sample data.Finally,a data set containing multi-target annotation information of motor vehicles,non-motor vehicles,pedestrians and some types of traffic signs on the road has been produced.The data set contains a total of 17,580 pieces of image data and annotated 138851 target instances.(2)Aiming at the problem of different target scales when detecting multiple types of targets at the same time,resulting in poor overall target detection results,this paper improves the clustering algorithm and selects a target detection setting frame that is more suitable for simultaneous detection of multiple types of targets.Drawing lessons from the idea of density clustering,this paper redesigned the distance formula of the K-means clustering algorithm.The design of the distance formula fully considers the impact of the density of the data near the cluster center on the clustering results.The density of the data near the cluster center is input into the distance formula as a similarity index to measure the effect of clustering,which solves the problem of the K-means clustering algorithm is susceptible to noise and the problem of poor clustering effect caused by uneven samples in the sample frame.Through target detection comparison experiments,it is found that the target detection setting frame obtained by the improved the K-means clustering algorithm can better reflect the distribution of target detection frames in target detection,and has good adaptability in target detection.And higher average accuracy of target detection is obtained..(3)Aiming at the problem of unbalanced samples when detecting multiple types of targets on the road at the same time,resulting in poor detection of small targets,this paper selects the single-stage target detection algorithm YOLO V3 as the research basis,starting with the loss function of the network model and the network structure diagram,improving the YOLO V3 target detection algorithm.First,the Focal Loss function is introduced to replace the prediction box positioning sum-of-square loss function,confidence cross-entropy loss function,and category probability cross-entropy loss function in the original YOLO V3 algorithm,which speeds up the initial convergence speed,reduces the effect of gradient disappearance,and improves the average accuracy of small target detection.Then,on the basis of the original YOLO V3 algorithm network model,the feature map obtained by 4 times downsampling and the 52×52 scale feature map convolution operation and the 2 times upsampling feature map are fused to obtain a higher fine-grained feature map.By adding a104×104 scale target detection module for small target detection,the neural network’s ability to detect small targets is improved.(4)Train and test the improved YOLO V3 algorithm on the self-made road multi-type target detection data set,and complete the comparison experiment of target detection before and after the improvement of YOLO V3 algorithm and the comparison experiment of the improved YOLO V3 algorithm and other target detection algorithms.Experimental results show that the improved YOLO V3 algorithm has higher average accuracy of multi-type target detection on the self-made data set than other target detection algorithms,with an average accuracy of 78.17%,which is 3.81 percentage points higher than the unimproved YOLO V3 algorithm.The speed has also reached 46.23 frames per second,which meets the requirements of real-time detection.