Research On Object Detection And Tracking Of Small Traffic Signs Based On Improved YOLOv3

Accurate and efficient road traffic signs detection is an important guarantee for the realization of the project landing of driverless vehicles.However,traffic signs are easy to be blocked,and affected by weather,lighting,and other external factors in the case of complex and changeable road conditions,resulting in uneven illumination,fuzzy,and a series of problems.In addition,the video of road traffic signs captured by the on-board equipment of the driverless car has a small visual range,how to detect and identify traffic signs quickly and accurately has gradually become one of the difficulties in self-driving tasks.Moreover,there are many kinds of traffic signs,and the frequency of traffic signs in different areas is different,and the distribution is extremely uneven.The rich variety and similar appearance of the characteristics also further increase the difficulty of road traffic signs detection at the same time.Therefore,the study of road traffic sign detection in the Real-World scenarios is of great significance.It is hard to balance the detection accuracy and real-time performance in practical application because of the small target of road traffic signs and the unbalanced distribution of all kinds of signs.To tackle this issue,this paper takes128 types of major traffic sign detection in China as an example,proposes an improved YOLOv3(You Only Look Once)model,and integrates the multi-object tracking algorithm Deep-Sort(Deep Simple Online and Real-time Tracking)to realize the detection and tracking of road traffic signs with small targets,multi-classification,high precision and high real-time in the case of severely unbalanced dataset.The main research work of this paper is as follows:1)Data augmentation.In view of the unbalanced frequency of different categories of traffic signs in the original dataset and the problem of under-fitting of the detection model caused by them: First,the traffic sign categories in the original dataset TT100 K are further subdivided,and the 128 traffic signs in the original dataset are extended to 152.Then,15 data enhancement methods,including noise,blur,rain,and snow corruption,are used to enlarge the uncommon traffic signs that appeared less frequently in the dataset,increasing the 6,605 images in the original dataset to 13,238 which contains over 20,000 traffic signs.As a result,the proportion of low-frequency traffic signs in the dataset increases,which improves the balance of sample distribution and overcomes the under-fitting caused by insufficient samples,thus improving the overall detection accuracy.2)The improvement of the YOLOv3 network.Aiming at solving the detection difficulties caused by small road traffic signs,the original YOLOv3 model structure is improved to make it more fit the detection requirements of small object detection.The detailed method is removing the output feature map corresponding to the 32-times subsampling of the input image in the original YOLOv3 structure to reduce its computational costs and adding a new output feature map of 4-times subsampling to improve its detection capability for the small traffic signs.Finally,to verify the effectiveness and the advantages,compared with state-of-the-art methods(e.g.Overfeat,MSA＿YOLOv3),our method shows better performances in detecting small target traffic signs,which obtains 91% precision,90% recall,and 84.76% m AP(mean average precision).3)Model pruning of the YOLOv3.In order to further improve the detection speed of the model while ensuring the accuracy of the model,the pruning operation is carried out for the proposed method in this paper.By measuring the importance of each parameter,the insignificant network layer is deleted,in this case,the structure of the model simplified and the parameters are reduced.Compared with Tiny-YOLOv3 and deep separable convolution methods,the effectiveness of the pruning method in this paper is verified from three aspects of model size,detection speed,and detection accuracy.4)The multi-object tracking algorithm is combined with the improved YOLOv3.Owing to the lighting and weather interference in the real traffic environment and the onboard camera motion blur,bumps,etc.in the video detection,resulting in missed detections and false detections.The safety of self-driving vehicles may probably be threatened.Hence,Deep-Sort is applied to object detection,which uses Mahalanobis distance,the smallest cosine distance,and other indicators to associate various targets in the video frames,improving the lack of tracking capability of the original YOLOv3 detection method.While stabilizing the actual video bounding box,the error rate and omission rate of video detection effectively decreased,and the anti-interference performance of the detection algorithm is enhanced as well.The comparative experiments in this paper show that :(1)It is demonstrated that the proper sample equalization method can effectively improve the performance of the model;(2)By improving the structure of YOLOv3,the real-time performance of video object detection can be improved on the premise of ensuring the detection accuracy of objects with different sizes;(3)The integration of Deep-Sort algorithm into the improved YOLOv3 improves the tracking ability of YOLOv3 and the anti-jamming ability of real-time object detection and tracking.The research results of this paper provide a technical reference for the self-driving system of driverless vehicles under real-scenes conditions,and have a broad prospect in a variety of autonomous driving applications such as unmanned road patrol and autonomous driving rental.