Research On Object Detection And Stability Strategies Based On Deep Learning

With the development of artificial intelligence technology,many emerging AI products have already landed.Among them,the highly-regarded autonomous driving technology has become more mature,and some vehicles with assisted driving functions have been put into production in batches.Object detection technology in the field of computer vision is one of the core technologies for realizing autonomous driving.It recognizes and detects 2D objects on the road to help drivers deal with road conditions in real time and avoid risks,thereby ensuring driving safety and reducing accidents.In the current real-time object detection process of autonomous driving,on the one hand,the accuracy of object detection needs to continue to improve,on the other hand,the problems of false detection and missed detection of detection objects still exist.During the driving of the vehicle,any obvious detection error may cause accidents,so it is necessary to improve the stability of object detection,thereby increasing the safety of vehicle driving.In view of the above problems,this article studies from three aspects.First,this paper is based on the deep learning framework pytorch,and uses the commonly used two-stage object detection framework R-FCN to train the autonomous driving data set to obtain the initial model.It is implemented and improved by combining the strategies for improving object detection performance in multiple papers.To further improve the performance of the model.Secondly,this paper proposes a strategy based on neural network attacks and defenses to improve the stability of object detection.The matrix box used to contain the size of the detected object in object detection is called the detection frame.The R-FCN detection framework Make improvements,add new network branch structures,and improve based on the FGSM(Fast Gradient Sign Method)method.Use the original data set to generate adversarial sample data,mix the adversarial samples and the original data for multiple strategy model training methods,and allow object detection more stable.As a result,flash frames in continuous video are reduced,and the number of missed and false detections in the detection frame is reduced,making the vehicle safer.Thirdly,this paper analyzes the main existing problems and proposes an index for evaluating the effectiveness of object detection stability.It comprehensively evaluates the number of flash frames,the center offset difference of the detection frame and the proportion change of the detection frame.The detection frame is tracked in the autopilot data set to obtain continuous frame detection frame information.Various strategies proposed in the article are separately trained and evaluated to find the best strategy to improve the stability of object detection..Through experiments on the autonomous driving dataset and the test set with the same distribution,the detection and evaluation indicators adopt the general evaluation indicator mAP(mean average precision),the basic model mAP is 0.798,and the improved effect is 0.844,which is an increase of 4.5 percentage points.After network stability,the flash frame was reduced by 30%,which verified the effectiveness of the strategy.The strategy of this paper improves the accuracy of object detection,increases the stability of the continuous frame object detection frame,and can be used in the actual automatic road object detection.