Research On Pedestrian Detection Algorithm Based On Dual Band Images

Pedestrian detection technology has a great application space in the robot,video security,intelligent transportation,intelligent driving field and so on.It has been received a lot of concerns from the researchers.Pedestrian detection technology can be divided into shallow machine learning method and deep learning method.The former algorithm mainly depends on the artificially designed features,and the latter will automatically extract and combine features from the data.Generally,pedestrian detection algorithms mostly focus on visible images,and some focus on infrared images.But visible images cannot afford enough good information in some complex conditions such as light changes,shadows,dim and other environments.Infrared images,due to the thermal imaging principle,are prone to heat interference.The details of infrared images will be seriously missed.Therefore,after the full investigation of pedestrian detection technology,this paper focuses on how to effectively use the visible and infrared images to improve the detection accuracy.The main work of this paper:(1)Compared with the single visible band,single infrared band and dual band pedestrian detection algorithm,we can find that the multi-sensor fusion method is more robust and adaptable to the scene change.Then,analyze and compare two pedestrian detection algorithms,that is aggregate channel feature based on shallow machine learning and faster region convolution neural network based on deep learning.(2)An improved pedestrian detection algorithm based on dual band aggregate channel feature is proposed.Firstly,the aggregate channel features of the visible image and the infrared image were extracted respectively.Next,different filter banks were designed to filter the dual band aggregate channel features.Finally,trained the classifier to realize the dual band pedestrian detection.(3)A pedestrian detection algorithm based on dual band convolution neural network is proposed.The algorithm uses a faster region with convolution neural network framework.Firstly,use a pre-trained model to fine-tune and optimize parameters.Then construct a proper dual band neural network by adding the batch normalization layer,the dual band concatenation and the dimension reduction operation,and the global average pooling layer.Experiments show that,compared with the result of single band images,dual band convolution neural network can do a better job in pedestrian detection.