Research And Design Of Key Technologies Of Vehicle Detection Based On CNN

According Statistics from the Ministry of Public Security show that the number of car ownership in China exceeded 240 million,and the number of motorists reached 369 million in 2018.With the rapid growth of the number of traffic vehicles,traffic supervision is facing a enormous challenge.Vehicle target detection,as a key technology for video surveillance of traffic conditions,has long been widely concerned by researchers at home and abroad.Traditional method of machine learning cannot accurately describe and cover the sample features since its ability of generalization is poor,making it difficult to achieve accurate recognition.The target detection algorithm based on deep learning features too large network structure and computational complexity,which can not meet the high real-time requirements of vehicle detection tasks.Therefore,it’s difficult to be applied in the actual vehicle detection.In view of the above problems,this paper aims to optimize the existing target detection model,as well as design and implement an algorithm verification platform to verify the improvement effect.The specific work is as follows:First of all,improve the network structure of the classic target detection model Tiny YOLO and propose a lightweight real-time one SPTNet.Optimize the model structure-Feature extraction component reduces the network redundancy weight to improve the model speed;Feature fusion component introduces the FSPP structure to better capture the multi-scale features of the target;Feature regression component uses the anchor boxes mechanism to make the candidate frame generated by the model more accurate and make the model easier to converge.Verify the proposed improvement of SPTNet-First,compare the results of SPTNet and Tiny YOLO models by using the public dataset Pascal VOC and the self-acquired vehicle dataset Anngic to verify the detection effect of SPTNet.Then,carry out the ablation experiments on each component of the model to verify the effectiveness of the proposed component optimization operations.Secondly,improve the algorithm of loss function,making the model more capable of learning samples with different levels of difficulty in data aggregation.In postprocessing,add some means to effectively improve the visual effect of detection.In this way,the real-time speed of the vehicle detection met the requirements,also the detection accuracy was improved.Robust experiments were performed by enhancing data of the data set to verify SPTNet’s ability to extract vehicle features.Experiments show that the detection effect is much better after the improvement.Finally,design the algorithm verification platform which aims to reduce the cumbersomeness of the model training process and provide the researchers with an easyto-use automation platform.It also offers a visual interface which is designed to enhance the operability for the developers.In addition to implementing the training and verification of the model,the platform can also evaluate the performance of the model to determine whether its accuracy and speed can meet the requirements.In summary,this topic is based on the research of CNN target detection algorithm optimization,and attempts to apply it to the field of vehicle detection.To make it meet the real-time requirements of vehicle detection and improve the detection accuracy.At the same time,the algorithm verification platform is designed and implemented with good research value and application value.