Car Detection And Parsing With Hierarchical And-Or Model

With the increasing seriousness of traffic jam and illegal parking, intelligent traffic surveillance systems have been widely gotten attention. As a critical technique in these surveillance systems, car detection and parsing have been the hot topic in the field of computer vision and intelligent transportation systems. In complicated traffic scenarios, occlusion has been the most difficult problem because of the frequently co-occurrences of car-to-car and person-to-car patterns. Besides, the large geometry and appearance variance of car parts also introduces many challenges.From the perspective of visual cognition and statistical modelling, this paper systematically study and analyze: 1) occlusion, multi-view, context relationships and intraclass variations in car detection; 2) part localization and status estimation in car parsing. Our method integrates implicit and explicit occlusion modelling, context modelling, spatial and temporal modelling by professional And-Or Graph(AOG). This dissertation presents our study from the following four aspects:We present a coupling-and-decoupling car detection model, which implicitly model occlusion, multi-view and car-to-car context information. Our work mainly focuses on occlusion handling in real car images. Different from the strategy of single car modelling, we believe, car-pair, as the basic unit of car-to-car occlusion, can be used to composite different occlusion patterns. Besides, a car-pair embeds more visual information than a single car, which can be used to train a more robust car detector. To represent these visual patterns, we propose a three-layer AOG to quantize various occlusions. Experiments on real occluded car datasets verify the superiority of our model.We present a novel car model based on occlusion simulation and self-adapted learning. In this work, we make use of the CAD graphical technology to simulate car-to-car occlusions in real life. In a CAD engine, by setting different camera viewpoints and CAD car models, the occlusion manifold of cars can be simulated. We manually annotate 17 semantic car parts, and according to the visibilities of car parts in each image, we could get an occlusion data matrix. By the algorithm of graph compression, the structure of the AOG model can be learned. To catch the appearance in real cars, we learn the parameters of AOG under LSSVM framework on real images. Experiments on a series of occluded car datasets verify the effectiveness of proposed AOG on occlusion modelling.We present a hierarchical And-Or model that integrates context and occlusion modelling. In this work, a unified framework is proposed to integrate context and occlusion information in car detection. Here, context denotes the relative positions of N-cars( N ?1). To reduce the model complexity and increase the model flexibility, we introduce a new AOG. The model structure is learned by mining discriminative context patterns and occlusion patterns. This AOG is still a directed acyclic graph(DAG), and dynamic programming(DP) algorithm can be used in inference. Parameters of this AOG is learned by the weak-label structural SVM(WLSSVM). In experiments, the proposed AOG achieves state-of-the-art performance on several challenging car datasets.We introduce a new problem in our vision community- car part status estimation and fluent recognition. Compare to car detection, our problem has more challenges in terms of the large geometry and appearance variance of car parts. And we also propose a spatialtemporal And-Or Graph(ST-AOG) to integrate car detection, car part localization, car part status estimation, and fluent recognition. Proposed ST-AOG is still a DAG, and we use LSSVM to learn its parameters. Experiments on a newly proposed car fluent dataset verified the effectiveness of our model.Our work focuses on occlusion handling and the learning of model structure under the background of car detection and parsing, and achieves some progress both in theory and application. Our proposed models and algorithms can provide some guidances and reference values to computer vision, pattern recognition and machine learning.