SEGMENTATION-BASED APPROACH FOR OBJECT DETECTION

Abstract

Translate Article

This study proposes a segmentation-based approach for object detection, developed for analyzing aquatic behavior in controlled laboratory environments. The research focuses on overcoming detection challenges in long-term video recordings of bullheads housed in enclosed aquariums, where sediment drift, background instability, and partial occlusions often confound traditional tracking techniques. To address these issues, an approach based on the improved SLIC Superpixel segmentation method was proposed. The basic SLIC method was modified to incorporate multi-layer contrast features and neighborhood-based pixel uniformity checks. The proposed approach includes the following stages: preprocessing, segmentation, clustering, and post-processing. The preprocessing stage includes bilateral and median filtering, contrast and brightness normalization, and optional image upscaling to improve clarity. Subsequent background subtraction and context-aware thresholding within segmented regions help eliminate false positives caused by floating debris and occluded contours. At the clustering stage, a refined distance metric is introduced to evaluate pixel coherence in a multilayered feature space, which include LAB components, subtraction results, and histogram-equalized grayscale representations, improving segmentation accuracy. Additionally, at the post-processing stage fragmented object blobs are merged to enhance spatial continuity. Empirical validation was conducted on a dataset of bullhead video frames recorded under realistic aquatic conditions. The approach based on the improved SLIC Superpixel segmentation method demonstrated an increase in object detection accuracy of more than 6% compared to the approach based on the basic SLIC method. The modularity and simplicity of the proposed approach allow it to be easily extended to other biological objects − in particular, for the behavioral analysis of rodents − without relying on deep neural networks or computationally intensive frameworks, making it suitable for tasks in ethology, neuroscience, and precision aquaculture. Further research will be devoted to implementing the approach in real-time and advanced trajectory analysis.