Abstract
Surveillance cameras are widely applied for indoor occupancy measurement and human movement perception, which benefit for building energy management and social security. To address the challenges of limited view angle of single camera as well as lacking of inter-camera collaboration, this study presents a non-overlapping multi-camera system to enlarge the surveillance area and devotes to retrieve the same person appeared from different camera views. The system is deployed in an office building and four-day videos are collected. By training a deep convolutional neural network, the proposed system first extracts the appearance feature embeddings of each personal image, which detected from different cameras, for similarity comparison. Then, a stochastic inter-camera transition matrix is associated with appearance feature for further improving the person re-identification ranking results. Finally, a noise-suppression explanation is given for analyzing the matching improvements. This paper expands the scope of indoor movement perception based on non-overlapping multiple cameras and improves the accuracy of pedestrian re-identification without introducing additional types of sensors.
1. Introduction
In large-scale indoor environments such as office complexes or shopping malls, individual surveillance cameras typically cover only a small fraction of the total area. Effective monitoring of personnel movement requires a network of cameras; however, these networks often feature non-overlapping fields of view. This creates significant "blind spots" where individuals disappear from one camera's view and reappear in another's.
The primary technical challenge in this context is Person Re-identification (Re-ID): the task of identifying whether a person captured by Camera A is the same individual seen later by Camera B. While traditional Re-ID methods rely heavily on visual appearance, they often struggle in office settings where individuals may wear similar attire, change poses, or move through varying lighting conditions.
2. Methodology
The proposed system addresses these challenges through a modular framework designed to bridge the gaps between non-overlapping views.
2.1 System Architecture
The framework consists of four primary modules: (1) a camera link module to establish spatial-temporal constraints, (2) a data preprocessing module for detection and feature extraction, (3) an instance matching module for Re-ID, and (4) an occupancy and transition pattern recognition module.
2.2 Stochastic Transition Matrix
A key innovation of this research is the integration of a stochastic inter-camera transition matrix. Rather than relying solely on visual similarity, the system incorporates the building's physical layout as a "prior." By analyzing historical movement data, the system learns the probability of transitions between specific camera pairs. The final similarity score is a fusion of visual feature embeddings (extracted via a ResNet-50 backbone) and these transition probabilities. If two images look alike but the transition between those two cameras is physically unlikely, the system suppresses that match.
2.3 Experimental Setup
The system was validated using a custom dataset collected from the 10th floor of the Liuqing Building at Tsinghua University. Five synchronized 4K UHD cameras were installed across a 1,200 $m^2$ area, covering corridors and room entrances.
3. Results and Discussion
The integration of spatial constraints served as an effective noise filter. Visual Re-ID often generates high similarity scores for different individuals with similar clothing—a phenomenon described as "noise." The transition matrix suppresses these false positives by penalizing geographically implausible matches.
Experimental results demonstrated consistent performance improvements:
- Accuracy Gains: The inclusion of the transition matrix led to an average Rank-1 accuracy boost of 5.69%, with peak improvements reaching 8.21% on specific days.
- Peak Performance: The system achieved a maximum Rank-1 accuracy of 83.94%.
- Noise Suppression: Heatmap analysis confirmed that spatial context effectively "cleans" the background noise of the matching algorithm, leading to much sharper and more distinct matches.
4. Conclusion
This study demonstrates that combining deep learning-based visual features with spatial-temporal constraints significantly enhances person re-identification in non-overlapping camera networks. This approach improves accuracy without the need for additional sensors or invasive facial recognition. Future work will focus on incorporating temporal distributions (travel time) to further refine matching and moving toward full trajectory restoration for continuous indoor movement mapping.
Summary updated