Object detection in video surveillance using deep learning approaches

Abstract

Object detection remains a challenging task in computer vision, not only because an object may appear at different scales with different non-rigid transformations, but also because of the intra-class variations and the variations caused by different viewpoints and illumination. Furthermore, some objects may be deformable, and the pose of an object appears differently in different scenarios. Object detection based on handcrafted features and conventional machine-learning methods has been researched for decades and has a proven performance with real-world applications. With the rise of deep learning, many classification problems have received a radical enhancement, in terms of accuracy, by using convolutional neural network (CNN), and this has inspired the use of CNN for detecting objects. In this thesis, we will first review some conventional object-detection approaches which use handcraft features and classical machine-learning algorithm, and then will conduct a survey on several CNN-based object-detection approaches. The region-based approaches use a coarsely trained CNN to extract class-agnostic region proposals, and then feed these proposals into a deep CNN for further classification and localization refinement. Proposal-free approaches have also been proposed, in which the CNN is used for both classification and localization. In our experiments, we find that different convolutional layers have different semantic meanings. A good localization requires feature with fine-grained details and high semantic features are better for classification. Therefore, we decouple the classification and localization from a single layer as majority of detector does. As a result, we proposed a fast vehicle detector, with a novel lateral convolutional network, which engages residual learning to enhance the localization but retains a high recall rate. The proposed detector can achieve relatively high accuracy and real-time detection for video surveillance. We have also designed an efficient traffic-sign classifier for an IEEE competition. Recently, we have also explored 3D object detection based on point clouds. A point cloud, unlike an image, has irregular data structure, which allows to give the same geometric representation with different permutation of the data. Unlike most of the literature, which projects the point cloud data into a Bird's eye view map, and applies a common image-based model, we have proposed a positional imbedded voxel feature encoding layer, which can learn the voxel representation by applying a simple convolutional layer. Experiments have shown that our proposed architecture has great potential for 3D object detection, with purely point cloud data.