Real-time Object Recognition And Tracking Based On Residual Neural Network

Recently,the problems of throwing waste from vehicle(TWV)have been attracted more and more attentions from the academic and social community.Since it is difficult to detect the TWV behavior in the real-time traffic surveillance videos,how to effectively solve the TWV problem has became an urgen problem to be solved for the traffic police and municipal departments.Although the research on Deep Learning has achieved many great achievements and has been applied in various applications(e.g.,vehicle license-plate recognition and traffic flow forecasting),there are many problems that has not been fully solved.Such as,how to timely and effectively detect the TWV behavior in the traffic domain is still a hard problem to be solved.In general,the traditional methods(e.g.,support vector machine)generally cannot achieve satisfactory performance in the sense of robustness and accuracy.As we all know,the power of deep learning system has greatly promoted the development of the artificial intelligence.Thus,in this study,we try to effectively detecte the TWV behavior in real-time traffic surveillance videos in a moderate time by using the typical deep learning system —residual neural network.In this study,we proposed a novel deep model to efficiently detect the TWV behavior by means of the real-time target recognition and the tracking technology based on the residual network,which is named Nov-Res Net-20.In addition,we applied the well-known you only look one(YOLO)model in the detection of the TWV behavior.Our contributions can be summarized as follows:1)Firstly,inspired by the famous VGG-16,we proposed a twenty-layer deep residual neural network(i.e.,Nov-Res Net-20)which consists of 6 convolutional layers and 7 residual layers.Secondly,since the open TWV dataset is inaccessible,we prepared a TWV dataset which is composed of 1000 samples with 224 × 224 size for evaluating the performance of the proposed Nov-Res Net-20.Lastly,the experimental results show that Nov-Res Net-20 achieved best testing accuracy(i.e.,94.30%)compared with the famous VGG-16,VGG-19,Res Net-50 and extreme learning machine(ELM).In addition,our experimental results also demonstrated that Nov-Res Net-20 had promising capability in the sense of detecting TWV behavior in traffic surveillance video.2)Although YOLO model has earned a big success in the real-time detection,to the best of our knowledge,there are no work uses YOLO model in the traffic TWV detection.Thus we applied YOLO model to detect the TWV behavior in traffic surveillance video for a better detection performance.Furthermore,we expanded the previous proposed TWV dataset from1000 to 3348 samples.Our experimental results on the expanded dataset reveal that YOLO model achieved satisfactory testing accuracy,recall and detection speed.