Electronic device capable of assigning tracklet feature vectors for multi-object tracking and its operating method

Precision Multi-Object Tracking Device Based on Color Histogram Feature Vector Updating

This technology is an electronic device for multi-object tracking that generates feature vectors from color histograms of tracklets detected in consecutive video frames. When the discrimination accuracy between tracklets is low, it recalculates and reassigns the feature vectors for those tracklets. It selects targets for updating by verifying the discrimination accuracy between tracklets and between tracklets and unassociated detection responses in two stages.

When tracking objects in CCTV or autonomous vehicle camera footage, tracklets—short movement paths—are typically linked based on color similarity. However, issues often arise where color feature accuracy drops due to video noise or occlusions by obstacles. This leads to incorrect linking of tracklets from different objects, causing tracking paths to be broken or swapped.

This technology collects additional color histograms from sample regions in past frames for tracklets with insufficient discriminative power and updates feature vectors using dimensionality reduction by learning a weight matrix via partial least squares. This allows for stable linking of tracklets of the same object even in the presence of occlusions or noise. It can be applied to intelligent CCTV monitoring, pedestrian and vehicle tracking for autonomous vehicles, and sports video analysis, ensuring continuous movement trajectories for each object even in crowded scenes.

‍

‍

Key Features:

• A feature vector generation unit that creates a feature vector using the color histogram numerical values of the t-th frame for each of the multiple tracklets detected in g frames
• A component that calculates a first discrimination accuracy, which is the degree to which tracklets are distinguished from each other based on the similarity between tracklet feature vectors, and checks whether it is below a first threshold
• A component that, if the first discrimination accuracy is below the threshold, calculates a second discrimination accuracy based on the similarity between each tracklet and unassociated detection responses and compares it with a second threshold
• A feature vector updating unit that obtains additional color histograms from the sample regions of tracklets with low second discrimination accuracy and updates the feature vectors using a weight matrix based on partial least squares

‍

This invention was developed with support from the Ministry of Science and ICT for video-based environmental guidance technology for the visually impaired and low-vision individuals.

‍