Research On Pedestrian And Vehicle Detection Based On Convolutional Neural Network

Detection technology of pedestrian and vehicle has become an important research direction in the field of computer vision,since it has been widely applied in many fields,such as video surveillance,self driving and intelligent robot.Based on the pedestrians and vehicles in life scenes as objects,and the performance of existing object detection algorithms has been analyzed comprehensively,in this thesis,a pedestrian and vehicle detection method based on convolutional neural network is researched.The main contents of this thesis are as follows:1.An object detection method based on interclass context feature is researched.Considering the low detection accuracy of the existing object detection methods on vehicle,we construct the interclass context feature according to the different prior information and apply it to the classification task of object detection algorithm by the way of the weighted ratio to correct the confidence value of different classes.By introducing the interclass context feature into the algorithm,the detection accuracy of object detection algorithm for all kinds of object is effectively improved,especially on vehicle.2.An object detection method based on channel pyramid feature is researched.Considering the situation that some objects are skipped over or some errors occur during the detection on the obscured or small objects using the existing object detection methods,we construct the channel pyramid feature by the way of cascading the receptive fields with different scales for different channels of the same feature map and apply it to feature extraction module in the object detection algorithm.The performance on obscured or small objects has been obviously improved.3.A pedestrian and vehicle detection method based on information fusion between channels of different feature maps is proposed.By analyzing the performance of existing object detection algorithm on the pedestrians and vehicles in life scenes and considering the improvement has been researched in the first two chapters,we introduce the interclass context feature and channel pyramid feature into the algorithm.In order to further improve the detection accuracy on the pedestrians and vehicles,we take into account the context information between the channels of different feature maps.Because the learning effect of convolutional neural network is highly correlated with the diversity of dataset,we build a dataset special for the detection of pedestrians and vehicles in life scenes,and validate the effectiveness of the proposed methods on this dataset.Experimental results show that the proposed methods can significantly improve the detection accuracy of pedestrian and vehicle in life scenes,and obviously improve the detection efficiency of existing algorithms for pedestrians and vehicles that are obscured or small.