Video Object Detection Method Based On Feature Enhancement

With the development of deep learning and convolutional neural network,video object detection tasks show great potential in practical applications,such as intelligent video surveillance and autonomous driving,which are widely concerned by academia and industry.In recent years,researchers have extended the advanced object detection framework to video field,but they are faced with two challenges in videos,that is,scale variation and deformation anomaly.However,most of the work rarely takes these two problems into account,resulting in the quality of the detection results have not met the actual needs.This paper focuses on these two challenges in videos,proposes two targeted modules to constructs a novel video object detection framework based on feature enhancement.The main work of this paper includes the following three parts:1)In order to solve scale variation in videos,scale-aware module is proposed for enhancement feature.The module is a lightweight network structure,which is mainly composed of a set of dilated convolutional layers with parameter constraint and a buffer convolution layer.This module can change the receptive field of convolution kernel only by controlling the dilated rate,so as to flexibly obtain the multi-scale information of the object and ensure the lower computation cost.2)For deformation anomaly in videos,the two branch RoIs(Regions of Interests)feature extraction module is proposed to solve it.This module mainly includes position-sensitive RoIs feature extraction branch and contexture-sensitive RoIs feature extraction branch,which extract RoIs features with object details and context information respectively.Then,the RoIs features generated by the two branches are fused by dot product.3)A novel video object detection framework based on feature enhancement is proposed to balance the detection accuracy and speed.In this paper,two feature enhancement modules are introduced into the advanced object detection framework.Optical flow network is used to effectively model the temporal information between video frames.And multi-frame aggregation strategy is used to further improve the deformation anomaly in videos.The experimental results show that the video object detection framework proposed in this paper achieves 77.9% m AP(mean Average Precision)on the Image Net VID dataset,which is 5.9% higher than the single frame detector,e.g.R-FCN.