Object Detection And Recognition For Traceability Video Based On Deep Learning

With the increasing number of cameras installed, as well as the increasing demand for smart cities and public security, the use of artificial video surveillance has been far from meeting people's needs. Deep Learning is one of the most important breakthroughs in the field of artificial intelligence in the past ten years. Nowadays, Convolutional Neural Networks (CNN) of deep learning achieve excellent results and have a wide range of applications in face recognition and pedestrian detection, image classification and so on.This thesis attempts to apply the deep learning technology to the field of Traceability Computing and other engineering fields. For the problem that the video information can not be processed in the traceability monitoring video, this thesis designs an intelligent video surveillance system to meet the practical application requirements using the off-line neural network models trained by the collected data set and the public data set, combining with the video collection and processing technology to complete the online detection, recognition and display of objects in video.The key points of the research are attached as following:1. Use the frame difference method to solve the problem which is very difficult to determine the key frames of the video, design a method of moving target detection by GMM and frame difference method based on the needs of target detection. The method can effectively overcome the foreground extraction process noise and shadow interference.2. Collect a set of target from traceability video and complete the target classification and annotation. Design and train the off-line convolutional neural network model based on this data set by using Caffe. The classification model has a higher accuracy than traditional method.3. According to the system demand, based on "motion detection+CNN" and "Faster-RCNN", use off-line convolutional neural network model, Redis and web application platform LAMP to design and implement the real-time target detection and recognition system.Finally, on the K40 Tesla hardware platform, the test accuracy of CNN model trained in the traceability video data set can reach 90.87%, the average processing time of "motion detection +CNN" system on one picture is 720ms, the average processing time "Faster-RCNN" system on one picture is 230ms. Test results show that the target detection and recognition system can meet the needs of the practical application of the functional and performance requirements.