Electronic device and operating method for assigning tracklet feature vectors for object tracking in video.

Tracklet Feature Vector Assignment for Multi-Object Tracking Using Random Weight Matrix Calibration

This technology is an electronic device and operating method that generates a feature vector by creating a color histogram vector for each tracklet detected in g frames and reducing its dimensionality using a random weight matrix. If the discrimination accuracy between tracklets or between tracklets and detection responses falls below a threshold, the weight matrix is calibrated to update the feature vector.

Tracking objects in CCTV or autonomous vehicle camera footage requires detecting tracklets—short movement paths—and connecting them across all frames. If the feature vectors assigned to these tracklets are not sufficiently distinct, different objects can be incorrectly linked, leading to inaccurate tracking paths.

This technology calculates a first discrimination accuracy between tracklets and a second between tracklets and detection responses using cosine similarity. For tracklets with low discriminability, it creates a calibration matrix using a weighted average of a new random matrix and the existing matrix to iteratively update the feature vectors. This allows for distinct feature vectors for each object without large-scale training. Applicable to intelligent CCTV, autonomous vehicle tracking, and retail customer path analysis, it reduces ID switching in crowded scenes and ensures accurate tracking.

‍

‍

Key Features:

• A histogram vector generator that creates a color histogram vector for each tracklet in the t-th frame when multiple tracklets are detected across g frames by an object tracking detector.
• A feature vector generator that randomly creates a first weight matrix with values between 0 and 1 and performs matrix operations to reduce the dimensionality of the histogram vector, thereby generating a feature vector for each tracklet.
• A calculation unit that computes a first discrimination accuracy based on feature vector similarity between tracklets and a second discrimination accuracy based on similarity with unrelated detection responses, then checks if they fall below a threshold.
• A feature vector update unit that generates a calibrated weight matrix by taking a weighted average of a second weight matrix and the first weight matrix for tracklets with a second discrimination accuracy below the threshold, updating the feature vector accordingly.

‍

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.

‍