SKFA-LSTM: Spatial Keypoint Feature Aggregation with Long Short term Memory for Predicting Road Traffic Accident

Abstract

Predicting road traffic accident concentrate on evaluating patterns in traffic data to identify potential risks. By employing significant factors such as vehicle movement, environmental conditions, and road interactions enables proactive decision-making, that reduce accident severity and improve overall road safety. However, capturing complex relationships in traffic data is challenging due to the involvement of highly dynamic, non-linear, and spatio-temporal dependencies, which leads to suboptimal performance. This research proposes Spatial Keypoint Feature Aggregation with Long Short-Term Memory (SKFA-LSTM) to predict road traffic accidents accurately. In LSTM, SKFA is included to select most informative spatial regions as intersections and vehicles. This improves the quality of input features and ensures that the LSTM learns temporal dependencies from significant patterns. Hence, the proposed SKFA-LSTM achieves a high mean Average Precision (mAP) of 93.00% and 99.98% on DOTA and CCD, compared to existing methods like TempoLearn..