Research On Vehicle Target Detection Method In Complex Weather Environment

With the development of social economy,there are more and more vehicles on the road,which brings great challenges to traffic supervision.Road traffic supervision mainly uses surveillance videos or images to detect vehicle targets.However,in actual traffic scenes,complex weather conditions often occur,such as rainy days,snowy days and foggy days,which will greatly reduce the imaging quality of cameras,thus affecting the accuracy of vehicle detection.Vehicle target detection is a key step in intelligent transportation,which is of great significance to follow-up vehicle tracking and vehicle type recognition.Therefore,it is necessary to study the restoration of complex weather images and the task of vehicle target detection,which has always been the focus of researchers.A rain removal algorithm based on multi-channel and multi-scale convolutional neural network is proposed.Firstly,the rain image is decomposed by guided filtering to get high-frequency detail image and low-frequency background image,in which the detail image mainly contains rain line texture,while the low-frequency image mainly contains background with high contour retention.Then,the detail image is filtered again to get a deep detail image,in which the rain line texture is more obvious,which can reduce the situation that the background texture in the detail image is misjudged as rain line texture when learning rain line features.Secondly,aiming at the residual raindrop imprint in the low-frequency image,it is proposed to send the low-frequency image and the deep detail image into the convolutional neural network for feature learning.In order to extract richer features of rain lines,multi-scale feature information is extracted from deep detail images by using hole convolution.Finally,the restored images with rain lines removed are obtained by combining the low-frequency images and deep detail images output by the network.In snowy environment,according to the imaging mechanism of snowflakes,snowflakes are in the form of white noise on the image surface,so they are regarded as salt and pepper noise.The method of image decomposition and re-filtering is used to eliminate snowflakes.Firstly,the snowcovered image is decomposed into detailed snowflake parts and low-frequency background parts by guided filtering,and then snowflakes are eliminated by bilateral filtering respectively.Finally,the detail parts and low-frequency parts after snowflake elimination are fused to obtain a de-snowed image.Compared with the de-snowing results of other single filtering methods,the snow removal effect of this method is the best.In foggy weather,the dark channel prior defogging method is used to defog the image,and the bilateral filtering algorithm is used to refine the coarse transmittance,so as to obtain a more detailed transmittance image,which provides better results for the final restoration of the haze image.This paper proposed a vehicle target detection algorithm based on improved YOLOv3 network to solve the problems.Such as,low efficiency,poor target detection effect and high missed detection rate of traditional vehicle target detection methods.In order to improve the efficiency of vehicle detection,a lightweight model MobileNet v2 is used to replace the feature extraction network in the original YOLOv3,which reduces the computational complexity of the network compared with the original algorithm.In order to effectively improve the detection ability of small-scale vehicle targets,the network uses the fusion of low-level feature map and high-level feature map for multi-scale detection.At the same time,aiming at the specific application of vehicle target detection,the anchor frame is re-clustered by K-means method to meet the needs of vehicle target detection.