Anticipation of Traffic Accident of Non-Ego Vehicles by Deep Learning Model.

Abstract

Safety is one of the most important issues that one should be aware of while driving a vehicle or in any type of autonomous vehicle. Human drivers should prevent and anticipate any kind of traffic accidents that may occur in any complicated condition. To prevent such kinds of accidents in the future, it should be anticipated as precisely and earliness of the accident. The early anticipation of non-ego-vehicle would caution the driver to take necessary action to avoid collision with non-ego-vehicle accidents. It would not only prevent an accident on the roads which result in costs to the car but also be deadly to the precious lives of the ego drivers. Recent research could anticipate accurately and early when accidents would occur for the task of accident anticipation. However, this research is only limited to ego-involved accidents. Therefore, the proposed model anticipates non-egoaccidents cases. This thesis focuses on the goal to improve traffic safety by anticipating traffic accidents on the road, based on deep learning. This thesis proposes a new deep learning approach that utilitzes the information from 1st person camera, also called dashcam video, to anticipate accidents related objects that may collide with non-ego vehicles. The proposed method predicts the accident probability of two other objects using two combinations of hidden states from a future object localization model which uses only motion features that are bounding box and its differential motion. The new method is based on four stages. The first stage is the detection and tracking of each object. This stage uses Mask-R CNN to detect and make a bounding box on each object on the road. Additionally, a deep sort algorithm is to be used to create an object id on each object. The second stage of the trajectory prediction part predicts the future location of the detected objects that are represented by bounding boxes and creates tracked IDs on each object. This stage generated hidden states. Further, the third stage consists of a combination of two hidden states of each object which are then matched with the annotation to check whether the combination of two objects collided with each other and then they are input to the next stage. The last stage is the accident anticipation part which computes the accident probability by using the combinations of two collided hidden states. The proposed method was validated on the dataset that includes multiple non-ego-accidents videos and normal driving videos. The proposed method is discussed with the previous method which was performed on ego-involve-vehicles, which show Average Precision (AP%) of 53.02 with an Average time-to-collision (ATTCs) of 0.69s.