The Identification of minor impact collisions in a long video for detecting property damages caused by fleeing vehicles using 3D convolutional neural network

Abstract

A parked vehicle damaged by a hit-and-run can be only repaired by the expense of the owner unless the fled vehicle is identified and the driver of the fled vehicle reprehended. Identifying the fled vehicle adopts a video investigation method that searches for perpetrators through recorded CCTV of the crime scene. When the length of the recorded video is long, the investigation may require extended amount of time of the investigator, resulting in an added burden to the investigator's daily work. Some commercial companies are using object recognition and tracking technology to detect hit-and-run, but the detection of small shaking of a vehicle during collision still remains as a challenge. Therefore, there is a need for a system that can detect small shaking of a vehicle in a lengthy video. The automatic recognition and tracking requires sufficient amount of training dataset. However, such dataset for hit-and-run is not publically available because it may violate personal information protection act. Rather than using the real accident video we may use actors to play such accident scenes. Although this may be feasible but such dataset would require substantial investments. In this paper, we describe a new dataset for hit-and-run. 833 hit-and-run videos were collected by recreating the parking lot using miniaturized cars. This dataset will be opened publically through Kaggle. 3D-CNN (Convolution Neural Network), which is frequently used in the field of action recognition, is used to detect small shaking of vehicles in a hit-and-run. In addition, the [degree of input] of the surrounding environment of the affected vehicle during a hit-and-run in which the reference vehicle is specified and the [length of the input frame] are adjusted to compare accuracy. As a result, the degree of input of the surrounding environment and the shorter the input frame, the higher the accuracy.