Research On Automatic Driving Target Detection And Recognition Based On YOLOv5

In recent years,with the continuous development of artificial intelligence technology and people’s yearning for intelligent life,artificial intelligence technology has been applied in every field of life.Among them,automatic driving technology is one of the most popular fields in the application of artificial intelligence technology in recent years.Target detection and recognition is the core of automatic driving technology of motor vehicles.It is an important way for motor vehicles to perceive the surrounding environment in the process of realizing automatic driving,and will have an important impact on the control and decision-making in the process of automatic driving of motor vehicles.Therefore,it is of great practical significance to study and utilize artificial intelligence technology to realize the target detection and recognition in the autonomous driving function of motor vehicles.In the target detection and identification application of YOLOv5 network in the automatic driving scenario,the problems of target density and target occlusion lead to the defects of false detection,leakage detection and poor detection accuracy of small-scale targets.Based on the YOLOv5 network,this paper conducted the following research and related improvements:Firstly,the attention mechanism was embedded in the backbone of the YOLOv5 network to improve the feature expression ability of the network.Compared with the introduction of different attention mechanisms,according to the experimental results,the CA attention mechanism had the best effect on the detection accuracy of the YOLOv5 network,which improved the average accuracy by 0.72%.Therefore,CA attention mechanism is introduced to improve the feature expression ability of YOLOv5 network model.Secondly,in order to improve the detection and recognition ability of the YOLOv5 model on the small scale target,the FPN structure of the original network is replaced by the cross-layer weighted cascade Path Aggregation network(WCALPAN).Based on the input and output nodes of the same scale,the learnable cross-layer weighted cascade is adopted to integrate the shallow semantic information into the deep semantic information.It makes the regression of the prediction box more accurate in the target localization task,and the learnable way also reduces the information loss caused by the fusion of deep semantic information and shallow information.At the same time,an up-sampling operation was added to the feature pyramid part of the YOLOv5 model,which expanded the receptive field of the feature map,further enlarged the features of the small target,and enriched the semantic information contained in it.Then,in view of the problems of false detection and missing detection caused by dense targets and occlusions,the NMS algorithm in the prediction part of YOLOv5 network was improved,and the DIOU-NMS algorithm was used to replace the original network’s NMS algorithm.The DIOU-NMS algorithm could effectively avoid mistakes when removing the prediction box,and effectively reducing the rate of missed detection and false detection of the model.Finally,based on the improved YOLOv5 network model,the classified network cascades EfficientNetV2.The improved YOLOv5 network is used to locate the target to be detected in the image,and the targeted information is entered into the network EfficientNetV2 for detailed classification,which not only reduces the classification loss of the model,but also further improves the insufficient detection accuracy of small-scale targets.Although the cascaded network increases the complexity of the model,the detection accuracy of the model is significantly improved,and the real-time performance also meets the requirements of the automatic driving scenario.In conclusion,after training and testing on the modified Udacity autonomous driving data set,the detection method proposed in this paper achieves a mAP of 93.58%and an FPS of 39,which not only has good target detection accuracy,but also ensures real-time detection,and has certain practical application capabilities.