The management of evidence for traffic accidents under surveillance video faces numerous challenges. Currently, traffic accidents evidence heavily relies on manual video review, and manually stored in third-party databases. This method of evidence storage cannot ensure the reliable source and trustworthy storage of evidence. Therefore, we propose a framework that combines deep learning with blockchain technology for traffic accident evidence management. It focuses on securely storing accident evidence and efficiently extracting accident information, providing a practical solution for traffic accident evidence management. In this framework, blockchain is combined with IPFS to overcome the drawbacks of centralized evidence storage, ensuring the security of evidence while alleviating the problem of limited blockchain storage capacity. Furthermore, a traffic accident evidence forensic model (YOLO-MBC) is constructed, which uses artificial intelligence to replace manual evidence collection. This not only ensures the reliability of evidence sources but also enhances the efficiency of evidence collection. Finally, experimental results demonstrate that the framework achieves secure evidence storage and the accuracy of the accident evidence forensic model (YOLO-MBC) reaches 92.4%, meeting the practical requirements of traffic accident evidence management.