Edge-Augmented HLAC and Gaussian Distribution-Based Weighted Feature Extraction for 1-ms Abnormal Detection System in Logistics

Ultra-low delay vision systems play an increasingly important role in fields such as logistics, among which the 1-ms box abnormal detection system is crucial. In recent studies on abnormal detection, deep learning-based feature extraction methods have rarely focused on cost-effective ultra-low delay systems, while rule-based feature extraction methods extract insufficient features. Achieving ultra-low delay detection based on the traditional method of combining multiple features is considered. This paper proposes the following: (A) edge-augmented HLAC patterns , which enable the feature extraction of more gradient information from texture-less boxes. Additionally, integrating the edge-augmented operator into the HLAC patterns does not increase the system’s inference time; (B) a Gaussian distribution-based weighted feature extraction method that emphasizes more discriminative features, thereby enhancing the system’s robustness across various environments. Experimental results demonstrate that the proposed abnormal detection system exhibits superior performance and has been successfully implemented on the field-programmable gate array (FPGA), achieving a delay of 0.937 ms.