By combining two complementary clues—camera geometry and visual appearance—researchers at Institute of Science Tokyo, Japan, developed a new approach for preserving identities across multiple cameras. The method uses epipolar geometry to identify spatially consistent candidate matches and appearance similarity to distinguish between possible identities. The approach can be integrated with existing single-camera tracking systems without requiring environment-specific retraining.
Multiple cameras can provide different views of the same scene, helping track people as they move through an environment. However, when a person is temporarily hidden by another person, an object, or other obstruction, a camera may lose track of them. When the person reappears, the tracking system may treat them as a new individual and assign a different ID to the same person, creating an identity switch. Maintaining the same identity across camera views therefore remains a key challenge in multi-camera tracking.
The solution lies in combining two complementary clues: physical geometry and visual appearance. The geometric relationship between cameras can narrow down where the same person could appear, while visual appearance can help confirm the match. Building on these complementary signals, a research team led by Professor Masayuki Tanaka and Professor Masatoshi Okutomi from the Department of Systems and Control Engineering at Institute of Science Tokyo (Science Tokyo), Japan, in collaboration with NEC Corporation, Japan, developed a new approach to multi-camera identity tracking.
The findings were made available online on August 3, 2026, and presented at the International Conference on Pattern Recognition (ICPR) 2026 , held in Lyon, France, between August 17–22, 2026.
The core of the approach relies on epipolar geometry, which describes the spatial relationship between the views captured by two cameras. When a person is observed by one camera, this geometry helps determine where the corresponding person could appear in another camera's image. The researchers use this relationship to calculate an epipolar distance and eliminate candidate matches that are unlikely to represent the same person.
The remaining candidates are then compared using appearance similarity. Visual features extracted from images of the person provide additional information to distinguish between people who could be matches identified through camera geometry. By using epipolar geometry to verify spatial consistency before applying appearance-based matching, the system can combine complementary information from the two cues when establishing cross-camera identities.
"By combining epipolar geometry with appearance similarity, we can use the geometric relationship between cameras to constrain possible matches and then use visual information to distinguish between them," notes Tanaka.
Rather than requiring a new tracking system, the method is designed to work with existing single-camera tracking systems. Each camera can independently detect and track people, producing tracklets that may become fragmented when a person is temporarily lost. Using the proposed approach, these tracklets can be associated across different cameras to determine which fragments likely belong to the same individual. A further advantage is that the cross-camera association does not require environment-specific retraining.
"Our approach uses the known geometric relationships between cameras together with pre-trained appearance features, so cross-camera track association does not require additional training for each environment," explains Tanaka.
To assess the effectiveness of their approach, the researchers evaluated it using two multi-camera tracking benchmarks, MMPTrack and CAMPUS, which contain video data from multiple camera views for assessing how well tracking systems maintain person identities.
On MMPTrack, their method achieved an average IDF1 score of 65.48 (a measure of how consistently a tracking system maintains the correct identities of people across frames), compared with 62.30 for MCTR, an existing multi-camera tracking method. It also achieved a score of 56.92 for Higher Order Tracking Accuracy (HOTA), a metric that assesses both how accurately people are detected and how consistently their identities are tracked. This was comparable to the HOTA of MCTR, which was 55.77.
On CAMPUS, the proposed method achieved an average IDF1 score of 47.37, compared with 44.72 for ByteTrack, an existing single-camera tracking method. The results also highlighted an important limitation: severe occlusion can cause people to be missed during detection, leaving no tracklet for the system to associate across cameras. This means that reliable multi-camera tracking depends not only on accurately matching observations, but also on consistently detecting people in the first place.
Improving both person detection and cross-camera association could help extend the approach to more crowded and challenging environments. The approach could potentially support applications where maintaining consistent identities across multiple viewpoints is important, including security monitoring, transportation management, facility operations, and pedestrian-flow analysis. Future work could focus on improving person detection and cross-camera track association under severe occlusion, potentially extending the approach to increasingly crowded and complex environments.
