This study investigates the use of computer vision and deep learning techniques to predict pedestrian–vehicle collision risk at road intersections, with the aim of improving traffic safety. A vision-based framework is developed by integrating YOLO-based object detection and motion tracking to analyze surveillance video data, while a Long Short-Term Memory (LSTM) network is employed for time-series prediction to evaluate potential collision risks. The research begins with image annotation and data preprocessing to enhance custom object detection performance, enabling accurate detection of pedestrians and vehicles as well as tracking their movement trajectories. Based on the extracted spatiotemporal features, an LSTM-based classification model is constructed to learn motion patterns and predict collision risk. To validate the proposed approach, a controlled small-ball collision experiment is first designed to simulate real-world traffic interactions. This experimental setup is used to train the model for collision risk prediction and to adjust model parameters through validation analysis. Subsequently, the optimized model is applied to real intersection surveillance footage for further evaluation. The results demonstrate that the proposed system can predict collision risk up to 2 seconds in advance, achieving an accuracy exceeding 73%, with potential for further improvement through continued optimization. In the future, this research may be integrated with intelligent transportation systems and vehicle-to-everything (V2X) technologies to enhance road safety and reduce traffic collision incidents.