Research And Application Of Video Object Detection Algorithm Based On Deep Learning

Research on computer vision technology based on deep learning has become an important research direction in the field of artificial intelligence.Object detection tasks for images or videos have occupied a pivotal position among many intelligent visual researches and applications.Aiming at the domestic and international researches on computer vision tasks,this thesis compares and analyzes the object detection mainstream algorithms and proposes the multi-scaled deformable convolutional neural network for object detection.Also based on the differences between video stream data and image data,it is proposed to adopt dynamic key frame scheduling strategy and combine with visual feature propagation and aggregation technology to carry out in-depth research on improving the comprehensive performance of video object detection.The specific research work mainly includes the following contents:1.The object detection mainstream algorithms based on deep learning model are compared and analyzed,and the RCNN series and other classical network models such as SSD and YOLO are also summarized.The multi-scaled deformable convolutional neural network is proposed to improve the performance of image obj ect detection.The use of deformable convolution operation increases the ability to identify objects which have undergone geometric deformation.The fusion of high-level and low-level features on multi-scaled feature maps can improve the object detection ability for small objects and dense objects.Experiments show that the improved network model adopted in this thesis keeps the computation efficiency and improves the accuracy of obj ect detection.2.Due to large amount of data,motion blur,occlusion,low pixel and other characteristics in video data,the dynamic key frame scheduling strategy based on optical flow field prediction is used to do preprocessing for video data in this thesis.Combining the sparse feature propagation for non-key frames realizes the real-time requirement and using the dense feature aggregation method for invalid key frames to contact the video stream context information to improve the accuracy of video object detection.3.Based on the above researches,the application and analysis system based on deep learning and object detection model in intelligent traffic scene is designed and implemented finally.The system includes preprocessing module for image and video data,object detection module for image data,object detection module for video data and some results analysis and statistical modules,etc.