Research On Vehicle Light Signal Detection And Recognition Based On Deep Learning And Attention Mechanism

With the increasing usage of cars,traffic safety has become one of the issues that need to be solved in the development of intelligent cities.According to traffic laws and regulations,there is a specific set of semantic meanings of vehicle light during the driving process.Vehicle light semantic information are a key factor to assist driving decisions during real driving scenario,and it is difficult for vehicles to recognize the light semantics of the vehicles ahead automatically in a practical complex environment.Therefore,the methods of vehicle light signal detection and recognition proposed in this paper are of research significance and application value.The researches of vehicle light signal detection and recognition based on deep learning and attention mechanism in this paper mainly has the following 4-fold contributions:(1)In view of the scarcity of the existing public datasets for vehicle tail-light signals detection and recognition tasks,we collect a large amount of driving video data from real urban environments,and then screen,truncate,extract video frames,annotate images.Finally,we have created the vehicle light signal dataset(VLS-Dataset)that is difficult to study.It is also provides a strong data support for verification of our vehicle light signal detection and recognition methods based on deep learning.(2)According to the characteristics of the vehicle tail-light signal detection and recognition task itself and the difficulty of the data set and other environmental disturbances,such as the unclear vehicle outline and the confused signal expression in low-light,this paper adopts the research ideas from coarse to fine and designs and builds a coarse/global attention(CA)module,which makes full use of the overall information of the original image,can roughly classify and locate the probable object.And we proves that our proposed method can effectively improve the detection network performance through mathematical theory derivation and experimental comparison on a universal data set.(3)As the task of vehicle tail-light signal detecting and recognizing is susceptible to reflections under strong-light,roadside street lights,other vehicle lights,and other light-emitting objects within the shooting angle of view,to address this issue,we propose and construct a fine/local attention(FA)module on the basis of our coarse attention module to further optimized the network,which is used to deeply learn the detailed semantic information of the tail-light of the vehicle and further locate the tail-light area,hence reduces the confusion between the lights semantics and achieves the best recognition effect of the vehiclelight signal.(4)Based on the general object detection network,we add these two embeddable and detachable attention modules designed above to improve the average accuracy of the detection and recognition.Then we use the deep learning network optimization method to accelerate the network calculation speed and finally improve a lot.An efficient end-to-end vehicle light signal detection and recognition system has been set up eventually to implement real-time vehicle light signal detection and recognition and early warning on the input driving video,which can greatly enhance driving intelligence and security in practical application scenarios.