Object Detection And Tracking In Driving Videos Based On Deep Learning

As important research areas of computer vision,object detection and tracking have always been the focus of research.And there are widely used in practical problems such as security monitoring and video analysis.Object detection based on the original manual feature ex-traction method has stagnated as early as 2010,so research on object tracking has also fallen into a bottleneck.With the increase of CPU and GPU computing power,related research on object detection and tracking has turned to application of deep learning.Feature extraction algorithms based on CNN keep appearing.Still,effect of applying deep learning technology to object tracking is not satisfactory compared with traditional correlation filtering methods,but it also means new opportunities and challenges.At the same time,the application of the above-mentioned technologies to intelligent traffic monitoring,driverless and other ap-plication fields is also currently a popular research.The application of object detection and tracking models based on deep learning for pedestrian and vehicle in driving videos will benefit the development of related technologies.Based on the above analysis,this thesis will explore object detection and tracking for driving vedio based on deep learning.While realizing object detection and tracking research based on deep learning,this thesis will explore the implementation and optimization of object de-tection and tracking for driving videos.The main content of this thesis could be summarized into the following:1.MobileNet is used instead of RexNeXt as backbone of Mask R-CNN,and this algo-rithm model is lightweight under the premise of ensuring detection accuracy to adapt to future road monitoring and intelligent driving needs to be transplanted into small devices such as automotive embedded.2.The channel purning method is used to optimize the object detection model.During the Mask R-CCN training and learning process,redundant neural network channels are trimmed to accelerate the model’s computing speed.3.The spatial regular thought of SRDCF is integrated into the deep learning-like object tracking algorithm SiamFC to eliminate the boundary effect in the model learning process.And add time regularity to accelerate the learning speed of the model.Through the above research,this thesis successfully established a Mask R-CNN object detec-tion model which take MobileNet as the backbone network,and optimized the model using the channel pruning method,which successfully reduced the model volume and accelerated the calculation speed.At the same time,this thesis successfully adds spatial regularity and temporal regularity to the SiamFC model which improves tracking accuracy and speed of model.