With the rapid proliferation of video event data recorders (VEDRs), video file data from VEDRs are often used as the primary evidence in many fields, such as law enforcement. In this paper, we propose a method for reconstructing corrupted video files and capturing key events recorded in the video file for use as valid evidence. The method first extracts image features from each video frame and constructs a multidimensional vector. Subsequently, dimension reduction of these vectors is performed for visualization in low-dimensional space. The proper sequence of the video frames is restored by using a curve fitting technique for the low-dimensional vectors. Then, we calculate the change in the slope of the curve-fitted model to detect key events in video files. The proposed method generates significant results not provided by existing file recovery techniques.