Research on Pedestrian Detection and Head Counting Algorithm Based on Computer Vision

This thesis focuses on the research of computer vision-based pedestrian detection and crowd counting algorithms, aiming to enhance their accuracy and robustness in practical scenarios. Firstly, for pedestrian detection, a more accurate and efficient framework is proposed by improving the Single Shot MultiBox Detection (SSD) algorithm. The optimized algorithm is validated on commonly used evaluation metrics and datasets, and is compared with other deep learning algorithms. Secondly, for crowd counting, the research is conducted based on the Multi-column Convolutional Neural Network (MCNN) algorithm. A more accurate method for crowd counting is proposed by improving the MCNN algorithm, effectively addressing the crowd counting challenges in different scenarios. Experimental results demonstrate superior performance of the improved algorithm on common evaluation metrics and datasets. Lastly, through comparative analysis of experiments, the strengths, weaknesses, and applicable scenarios of pedestrian detection and crowd counting algorithms are summarized. The research outcomes of this thesis significantly contribute to improving the effectiveness of pedestrian detection and crowd counting algorithms in practical applications.