Research On Pedestrian Detection Based On Deep Learning

As an important content in the field of object detection, pedestrian detection have been used in many aspects of life,such as advanced driver assistant system,safety monitoring,intelligent robots and so on. Therefore,pedestrian detection is of great significance to pedestrian safety and even national security, and its research has good theoretical significance and great application value.The current pedestrian detection does not go deep into what kind of deep convolutional neural network, or what kind of deep learning method is suitable for pedestrian detection, this paper analyzes the network structure for pedestrian detection by means of conducting multiple experiments, and summarizes the tips that should be paid attention to in terms of data collecting and training. On the other hand, people only utilizes the feature of the last layer in the network, while ignoring the characteristics of the network middle layer. This paper proposes to fuse the features in different layers of network to obtain a better expression of pedestrian.For the purpose of exploring the performance of different networks, we implemented different nets based on the Faster R-CNN framework with Caltech and PSDB database. By constructing extensive experiments through altering the database, changing the quantity of the train data and comparing the detection rate of each training stage, we compared their generalization ability, learning capability and convergence rate of three deep architectures. Also, we chooses ResNet with the characteristics of conducting feature between different layers of network, fully integrates the features of shallow layer with deep layer. Based on PSDB, the detection rate was further improved.In this paper, we constructed extensive experiments on three kinds of classical network,systematically analyzed the methods of improving pedestrian detection, and got some conclusions:(1) When the data is exceeded 100 thousands, network of more than 8 layers is conducive to the performance. To a certain extent, deepen the network can improve the detection performance. When data is less than 30,000, the convolution network model is not recommended to select.(2) When layers of network is less than 8, the iterations should not exceed 50,000, otherwise over-fitting will happen.(3) Select the data-set containing plenty scenes is beneficial to the learning of network, PSDB is worth popularizing.These conclusions is helpful for the follow-up pedestrian detection based on deep learning. Also,the proposed method of fusing the features of different layers has a 2.1 % lower than that of VGGNet in miss rate, which shows that the fusion of feature in different layers can improve the performance, this conclusion provided some reference for relevant work.