Non-motor Vehicle Detection And Recognition Based On Deep Learning

In recent years,with the rapid development of new business formats such as takeaway,express delivery services,and shared travel,the surge in the number of non-motor vehicles has brought huge difficulties to road traffic supervision.Non-motor vehicles have the characteristics of small object,large number,and relatively flexible movement which make the method of traditional object detection stretched.Therefore,research on the detection and recognition of non-motor vehicles based on convolutional neural networks in the context of deep learning can have an important positive influence on the development of intelligent transportation systems,and the improvement of technologies such as unmanned driving.We studies the positioning and recognition process based on the Faster R-CNN algorithm,analyzes the object characteristics of non-motor vehicles,and combines the three aspects that positioning effect,training strategy and feature network optimization algorithm to comprehensively improve the accuracy and reliable of non-motor vehicles detection and recognition,the main work of this paper is as follows:(1)It founds that the RPN of Faster R-CNN has the defect that the loss function cannot reflect the actual overlap in the coarse positioning.Consider using IoU directly as the loss function and using the improved version of the IoU loss function to improve this situation,the effect is not satisfactory.we uses the improved MIoU loss function,which effectively improves the accuracy and speed of the object positioning regression.(2)In the object recognition stage,the original VGG backbone network only extracts and utilizes the deep convolutional layer.Although the deep feature map contains rich high-level semantic information,it lacks the coverage of detailed information such as object contour and texture in the shallow feature map.we used an algorithm of object fusion enhancement based on feature fusion,which realizes the efficient combination between deep and shallow feature information,and effectively improves the information expression ability of the feature map,the classification information of the object is enriched.(3)The IoU threshold of the RPN in the training phase will causes a mismatch in the regression effect of the proposals between the test and training phases,making it impossible to identify the accurate pixel area in the classification phase.In this article,the cascade structure is used to redesign and expand the RPN.By setting the gradient threshold,it is realized that the continuous optimization of the corresponding low-quality proposals at each stage.The final RPN outputs the optimal object bounding boxes,which provides reliability for subsequent classification.(4)Combine the above improved algorithms,the complete algorithm in this article is given.System experiments and tests are carried out on the self-made data set.The average accuracy of the three types of non-motor vehicle objects has been improved.Finally,compared with the baseline network,the value of m AP(mean Average Precision)is increased by about 12%in this article,indicating that the proposed algorithm in this paper significantly improve the algorithm effect on the self-made non-motor vehicle data set and has high robustness.