Research on Pedestrian Detection Based on Improved SVDD Algorithm

Aiming at the problem that the SVDD target detection algorithm does not consider the impact of the probability density distribution of samples on its performance, and the detection accuracy is low, an improved SVDD pedestrian detection algorithm is proposed. Firstly, the HOG algorithm is used to extract pedestrian features and obtain the HOG feature vector. Then, the DHSDD model is designed to reduce the possibility that normal samples are located outside the hypersphere, and probability density estimation is introduced into the DHSDD model to obtain the DWDHSDD model. Replacing the probability density function in the DWDHSDD model with a membership function to reduce misclassification of noisy normal samples, thus constructing a pedestrian detection algorithm based on IDWDHSDD. The experimental results show that for pedestrian detection in different scenarios, the average accuracy is 94.05%, and the average missed detection rate is 6.05%; The precision, recall, and average precision of IDWDHSDD algorithm are higher than those of DWDHSDD algorithm, DHSDD algorithm, DSVDD algorithm, and SVDD algorithm, with 97.43%, 96.78%, and 97.14% respectively. Moreover, its precision, recall, and average precision have been improved by 12.27%, 11.69%, and 13.27% compared to SVDD algorithm; Its miss detection rate is 4.41%, which is much lower than the other four algorithms. Therefore, the improved algorithm in this article has high detection accuracy, low miss rate, good robustness, and excellent detection performance, which can provide technical support for pedestrian detection.