Research On Traffic Signal Detection Algorithm Based On Deep Learning In Complex Environment

With the rapid development of China's economy,car ownership has been increasing.However,it also causes frequent traffic accidents,and the number of casualties caused by traffic accidents is increasing year by year.In the face of increasingly complex road traffic conditions,the technology of vehicle-assisted driving,danger warning,pedestrian detection as well as the parallel detection which are based on computer vision technology has emerged.The traffic signal light identification is an important link in vehicle-assisted driving and vehicle hazard warning,and it also becomes the focus and hotspot of current research.Besides,recognition of traffic lights is a typical vertical application scenario in the field of target detection,it not only contains the target detection and recognition,but also has a very high requirement on its performance.The traditional target detection is divided into two steps.First,the candidate box of the target may be determined,then identify the target box.The selection of candidate boxes generally adopts the method of window sliding which has the following deficiencies.First,the number of candidate frames generated is huge,resulting in a relatively long detection time.Second,fixed candidate frame size that cannot bring accurate target background information.Third,traditional feature extraction requires the design of feature models.Manually-designed models tend to result in single features and the operation is very complex.It is difficult to obtain ideal results in practical applications.In recent years,the end-to-end feature extraction method based on deep learning has provided a new idea for the detection of traffic lights.Convolutional neural networks in deep learning are widely used in target detection due to their excellent feature generalization and dimensionality reduction features.However,there are still great challenges in the detection and identification of small objects,especially in the detection and recognition task of traffic lights,there are a lot of problems caused by the shooting distance.At the same time,self-illuminating light sources like car taillights will also affect the detection.This topic mainly analyzes the special task scenario of traffic signal detection and recognition,and the former researchers have ignored the influence about the background information on its detection rate of recall and recognition accuracy in the previous research.Therefore,this article is starting from the context information of the target,then carry out a further research.Based on the analysis of the error detection results in the existing algorithm,this paper also obtains the correlation between the different context dimensions and the test results in the target detection.Combining the actual size of traffic signal lamp,this topic put forward a method by using three different sizes of context information,it aims to improve its accuracy and recall rate by having a traffic signal detection and recognition assistance.This method has a significant advantage of improving the rate of traffic signal detection and recognition accuracy,but there are also existing some shortcomings,such as the redundant information,rigid position and so on.This article is putting forward a kind of context size calculation model based on image feature,then it can not only solve the problems that I have mentioned above,but also enhance the rate of recall and accuracy effectively which are compared with fixed size of the context.