Abstract:In case surveillance videos, slow moving small targets are prone to motion blur due to prolonged exposure or camera shake, resulting in feature loss or deformation, making it difficult to accurately capture their key features and reducing detection accuracy. Therefore, an improved OMP algorithm is proposed for enhancing the detection of slow moving small targets in multiple cameras. To address issues such as noise interference, Gaussian, median filtering, and Retinex algorithms are combined for denoising and enhancement, providing high-quality images for feature extraction of small targets. Using Faster R-CNN for small target feature extraction, capturing local features through convolution and pooling operations, generating candidate regions through RPN network, proposing a new confidence formula to accurately determine regions, and then obtaining feature vectors through RoI Pooling. Applying improved OMP algorithm to enhance small target features, introducing local correlation and energy distribution, combined with nonlinear transformation, to improve the enhancement effect and stability. The process of multi camera small target association and information fusion involves associating small targets based on feature vector similarity, using weighted fusion to process information, and forming a complete motion trajectory. Based on Kalman filtering and support vector machine, slow motion small target tracking and detection result output is achieved. Firstly, state estimation and covariance matrix are initialized, and continuous tracking is carried out through state prediction and observation updates. Then, SVM is input to determine the category. At the same time, visualization technology is used to annotate and display the results, including drawing external rectangular boxes, annotating position and category information, displaying video frames, etc., to provide users with accurate and clear slow motion small target information. The experimental results show that after applying the design method, the maximum contrast value of the video frame image reached 1.9, and the minimum error of small target feature reconstruction reached 0.4%, indicating high detection accuracy.
