Pedestrian Detection Based On Improved Faster RCNN Algorithm

With the development of deep learning,the field of object detection has attracted more and more researchers in image processing.Pedestrian detection is an important task in computer vision and pattern recognition,and has a wide range of application,such as autonomous driving,intelligent monitoring,and intelligent robots.The domestic and foreign research status of classic convolutional neural networks in recent years and the design criteria and optimization methods of convolutional neural network models are studied.In recent years,the use of deep learning algorithms for pedestrian detection can be tested on many public pedestrian datasets.Compared with the traditional machine learning algorithms and other image processing algorithms,the experimental results have achieved better performance.In videos and images,factors such as distance,insufficient light,shadow interference,or some pedestrians have smaller pixels,which makes it difficult for some pedestrian detection methods to detect such small pixel pedestrian targets,which increases the false detection rate.According to the above questions,this paper uses the improved Faster RCNN algorithm to detect pedestrians.The main research contains the following parts:1.In order to have a more intuitive observation of the output features of the convolutional neural network,we visualized the feature map and found that the shallow network extracted texture and detail features,which contained more features,and the deep network extracted contour shapes and highlighting features,the extracted features are more representative,but the fine-grained information of the feature map will be reduced.In this paper,the features of different convolutional layers are fused,and the extracted features combine the characteristics of low-level information and high-level information to describe the pedestrian characteristics of small targets.The experimental results obtain the miss rate of 10.31% on INRIA dataset,and the miss rate of 24.27% on the Caltech dataset.2.For the target detection task,the performance of the network greatly affects the detection effect.Compared with original Faster RCNN network,we combine with the SENet(Squeeze-and-Excitation Networks)to obtain the importance of each feature channel,and promote useful features and suppress features that are not very useful for the current task.Consequently,the performance of the network is improved through the relationship between the channels.The complexity of this algorithm is not high.The experimental results obtain the miss rate of 10.20% on INRIA dataset,and the miss rate of 13.14% on the Caltech dataset,and this method takes 0.28 s to detect an image on the Caltech dataset.