Intelligent Traffic Pedestrian Detection by Integrating YOLOv4 and Improved LSTM

Abstract

As an important component of modern urban traffic management, intelligent transportation systems aim to improve road safety and reduce traffic congestion. The study proposes an intelligent traffic pedestrian detection model that integrates YOLOv4 and an improved time recurrent neural network. The model utilizes attention mechanism to optimize YOLOv tiny for identifying human key points, and introduces multi-head attention mechanism to optimize the time recurrent neural network for pedestrian intention recognition. The performance test results showed that when only using pedestrian position information, the model accuracy was 53.2%, and the processing speed reached 76 frames per second. When relying solely on human keypoint information, the accuracy significantly improved to 72.6%, but the processing speed decreased to 20 frames per second. When combining pedestrian position and human key point information simultaneously, the model achieved the highest accuracy of 78.8%. The experimental results have proven that the intelligent traffic pedestrian detection system designed in the study can meet the real-time and accuracy requirements of vehicle mounted systems for pedestrian detection and can achieve accurate capture of pedestrian intentions.