Modelling active bio-inspired object recognition in autonomous mobile agents

Doctoral thesis English OPEN
Bermudez Contreras, Edgar Josue (2010)
  • Subject: QC | TJ

Object recognition is arguably one of the main tasks carried out by the visual cortex. This task has been studied for decades and is one of the main topics being investigated in the computer vision field. While vertebrates perform this task with exceptional reliability and in very short amounts of time, the visual processes involved are still not completely understood. Considering the desirable properties of the visual systems in nature, many models have been proposed to not only match their performance in object recognition tasks, but also to study and understand the object recognition processes in the brain. One important point most of the classical models have failed to consider when modelling object recognition is the fact that all the visual systems in nature are active. Active object recognition opens different perspectives in contrast with the classical isolated way of modelling neural processes such as the exploitation of the body to aid the perceptual processes. Biologically inspired models are a good alternative to study embodied object recognition since animals are a working example that demonstrates that object recognition can be performed with great efficiency in an active manner. In this thesis I study biologically inspired models for object recognition from an active perspective. I demonstrate that by considering the problem of object recognition from this perspective, the computational complexity present in some of the classical models of object recognition can be reduced. In particular, chapter 3 compares a simple V1-like model (RBF model) with a complex hierarchical model (HMAX model) under certain conditions which make the RBF model perform as the HMAX model when using a simple attentional mechanism. Additionally, I compare the RBF and HMAX model with some other visual systems using well-known object libraries. This comparison demonstrates that the performance of the implementations of the RBF and HMAX models employed in this thesis is similar to the performance of other state-of-the-art visual systems. In chapter 4, I study the role of sensors in the neural dynamics of controllers and the behaviour of simulated agents. I also show how to employ an Evolutionary Robotics approach to study autonomous mobile agents performing visually guided tasks. In addition, in chapter 5 I investigate whether the variation in the visual information, which is determined by simple movements of an agent, can impact the performance of the RBF and HMAX models. In chapter 6 I investigate the impact of several movement strategies in the recognition performance of the models. In particular I study the impact of the variation in visual information using different movement strategies to collect training views. In addition, I show that temporal information can be exploited to improve the object recognition performance using movement strategies. In chapter 7 experiments to study the exploitation of movement and temporal information are carried out in a real world scenario using a robot. These experiments validate the results obtained in simulations in the previous chapters. Finally, in chapter 8 I show that by exploiting regularities in the visual input imposed by movement in the selection of training views, the complexity of the RBF model can be reduced in a real robot. The approach of this work proposes to gradually increase the complexity of the processes involved in active object recognition, from studying the role of moving the focus of attention while comparing object recognition models in static tasks, to analysing the exploitation of an active approach in the selection of training views for a object recognition task in a real world robot.
  • References (65)
    65 references, page 1 of 7

    1 Introduction 24 1.1 Structure overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

    2 Active object recognition and autonomous mobile robots 27 2.1 Object recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.2 Object recognition models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.2.1 Computer vision approaches . . . . . . . . . . . . . . . . . . . . . . . 29 2.2.2 Biologically inspired approaches . . . . . . . . . . . . . . . . . . . . 30 2.3 Active perception, embodiment, and situatedness . . . . . . . . . . . . . . . 34 2.3.1 Active vision and object recognition . . . . . . . . . . . . . . . . . . 34 2.3.2 Embodied and Situated visual systems . . . . . . . . . . . . . . . . . 35 2.3.3 Movement and object recognition . . . . . . . . . . . . . . . . . . . . 36 2.3.4 Temporal information and object recognition . . . . . . . . . . . . . 37 2.4 Controllers for autonomous visually guided mobile robots . . . . . . . . . . 37

    3 A first comparison: HMAX and RBF models in realistic conditions 39 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.2 Visual system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.2.1 The Analysis module . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.2.2 Classifier module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.2.3 The attentional and foveation mechanisms . . . . . . . . . . . . . . . 49 3.3 Model evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.3.1 State of the art comparison . . . . . . . . . . . . . . . . . . . . . . . 51 3.3.2 HMAX implementation validation . . . . . . . . . . . . . . . . . . . 52 3.4 Comparison of the models in more realistic conditions . . . . . . . . . . . . 56 3.4.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

    5 Active acquisition of visual information 82 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 5.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 5.2.1 Agent, arena and objects . . . . . . . . . . . . . . . . . . . . . . . . 83 5.2.2 Training phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 5.2.3 Testing phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 5.3.1 Similarity maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 5.3.2 Testing the models using movement trajectories . . . . . . . . . . . . 94 5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 5.4.1 Dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 5.4.2 The role of the BDM . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

    6 Movement strategies during learning 104 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 6.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 6.3 Experiment 1: Movement strategies . . . . . . . . . . . . . . . . . . . . . . . 107 6.4 Experiment 2: Temporal information using the RBF model . . . . . . . . . 111 6.5 Experiment 3: Robustness of the RBF when using temporal information . . 114 6.5.1 Changing the radius of strategy 3 . . . . . . . . . . . . . . . . . . . . 114 6.5.2 Moving the centre of strategy 3 . . . . . . . . . . . . . . . . . . . . . 115 6.5.3 Using strategy 3 for training and the testing trajectory for testing. . 116 6.5.4 Moving the centre of strategy 4 . . . . . . . . . . . . . . . . . . . . . 116 6.5.5 Considering interval timing for strategy 4 . . . . . . . . . . . . . . . 118 6.6 Experiment 4: Using more objects . . . . . . . . . . . . . . . . . . . . . . . 119 6.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 6.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 7.4.1 Differences between the real world and the simulated case . . . . . . 137 7.4.2 Exploitation of variation in the object views in the real world . . . . 137

    7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

    8 Towards active selection of training views 139 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 8.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 8.2.1 RBF versions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 8.2.2 The classifier module . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 8.2.3 Movement strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 8.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 8.3.1 Reducing the complexity of RBF . . . . . . . . . . . . . . . . . . . . 144 8.3.2 Investigation of training views and model performance . . . . . . . . 146 8.3.3 Exploiting regularities in the environment through movement . . . . 151 8.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 8.4.1 How could the reduced versions of the RBF model fully regain performance? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 8.4.2 Towards active object recognition . . . . . . . . . . . . . . . . . . . . 160 8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

    1.1 Brighton seafront on a Sunday morning. . . . . . . . . . . . . . . . . . . . . . . 24

    2.1 Ventral and dorsal pathways in the visual cortex. The activity of the ventral pathway is generally associated with the identification of objects while the dorsal pathway is commonly associated with the localisation and actions related to objects in space (image adapted from wikipedia). . . . . . . . . . . . . . . . . . . . . . 31

    2.2 Hierarchical structure in the HMAX model. (adapted from (Riesenhuber and Poggio, 1999b)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

  • Related Research Results (1)
  • Metrics
    0
    views in OpenAIRE
    0
    views in local repository
    5
    downloads in local repository

    The information is available from the following content providers:

    From Number Of Views Number Of Downloads
    Sussex Research Online - IRUS-UK 0 5
Share - Bookmark