Abstract
The output of the residual network fluctuates greatly with the change of the weight parameters, which greatly affects the performance of the residual network. For dealing with this problem, an improved residual network is proposed. Based on the classical residual network, batch normalization, adaptive $\beta$-dropout random deactivation function and a new loss function are added into the proposed model. Batch normalization is applied to avoid vanishing/exploding gradients. $\beta$-dropout is applied to increase the stability of the model, which we select different dropout method adaptively by adjusting parameter. The new loss function is composed by cross entropy loss function and center loss function to enhance the inter class dispersion and intra class aggregation. The proposed model is applied to the indoor positioning of mobile robot in the factory environment. The experimental results show that the algorithm can achieve high indoor positioning accuracy under the premise of small training dataset. In the real-time positioning experiment, the accuracy can reach 95.37%.
Introduction
Indoor localization is a critical challenge for mobile robotics, particularly in complex industrial environments like factories. While traditional methods such as Wi-Fi, Bluetooth, or RFID are commonly deployed, they often suffer from signal interference, multi-path effects, and the requirement for expensive infrastructure. Vision-based localization using standard RGB cameras offers a cost-effective alternative.
However, vision-based systems face two primary hurdles: model instability and visual similarity. Standard Deep Residual Networks (ResNet) can be highly sensitive to small fluctuations in weight parameters, leading to inconsistent outputs. Furthermore, in indoor environments, adjacent locations often look nearly identical, making it difficult for standard classification models to distinguish between them. This paper proposes an "Improved ResNet" architecture specifically optimized to address these challenges.
Methodology
The proposed architecture enhances the standard ResNet-50 with three primary modifications: Batch Normalization, Adaptive $\beta$-dropout, and a Hybrid Loss Function.
Batch Normalization (BN)
BN is integrated to regulate inputs across layers, effectively preventing the vanishing or exploding gradient problems that often plague deep networks. This ensures more stable training and faster convergence.
Adaptive $\beta$-dropout
Unlike standard dropout mechanisms that utilize a fixed probability, the proposed adaptive $\beta$-dropout employs a Beta distribution to dynamically adjust the deactivation strategy throughout the training process:
- Early stage: A Bernoulli distribution ($\beta=0.001$) is used to prune unimportant nodes.
- Middle stage: A uniform distribution ($\beta=1$) is applied to smooth node activations.
- Late stage: A Gaussian distribution ($\beta=3$) is used to highlight the most critical features.
Hybrid Loss Function
To improve the discriminative power of the model, a hybrid loss function is introduced. It combines standard Cross-Entropy Loss with Center Loss. While cross-entropy handles general classification, center loss minimizes the distance between a sample and its class center in the feature space. This dual approach forces the model to create "tight" clusters for each location, significantly enhancing the ability to distinguish between visually similar regions.
Experimental Results
The Improved ResNet was evaluated on both the ImageNet 2012 benchmark and a custom factory dataset.
General Performance (ImageNet)
On the ImageNet dataset, the Improved ResNet-50 achieved a Top-1 error rate of 20.35%, outperforming the standard ResNet (20.74%) as well as Inception and VGG-16 architectures. This confirms the general robustness of the proposed enhancements.
Factory Localization Performance
The primary evaluation was conducted on a specialized factory dataset consisting of 1,800 labeled images across 18 distinct regions. Each location was represented by 100 images captured from various angles.
The improved model reached a positioning accuracy of 95.37%. This represents a 2.04% improvement over the standard ResNet implementation. Notably, the model significantly reduced errors in "middle regions" where visual features are most ambiguous. The confusion matrix in Figure 6 demonstrates the high precision across all 18 classes.
Conclusion
The Improved ResNet architecture demonstrates that integrating dynamic regularization and hybrid loss functions can significantly enhance the performance of vision-based indoor localization. By focusing on intra-class aggregation through center loss and adaptive stability through $\beta$-dropout, the model achieves high accuracy even with limited training data. Future work will focus on scaling the dataset to larger industrial plants and exploring the model's robustness under varying environmental conditions such as lighting changes and dynamic obstacles.
Summary updated