Crash prediction from camera images

As autonomous vehicle technologies advance, the need for reliable perception and response mechanisms to ensure safety in dynamic traffic environments is becoming critical. This work presents a novel hybrid model, combining R-CNN, a Kalman Filter, a Hidden Markov Model (HMM), and an Artificial Neural Network (ANN), to detect and predict vehicle collisions in real-time using a single camera input. The system performs object detection and multi-object tracking in video streams, applies probabilistic motion prediction via a modified Kalman Filter, and enhances state transitions using a hybrid approach of the HMM and ANN. The proposed architecture significantly increases detection accuracy and reduces latency under rapid motion, addressing the limitations of traditional filtering techniques. The model was tested in the CARLA simulation, on video footage, and in real-time. The conducted experiments demonstrate high efficacy, achieving 94.5% detection accuracy in video analysis and 91.2% accuracy in the real-time application. The ANN integration resulted in a 0.5-second reduction in detection latency, a critical improvement when considering average human reaction times. These results suggest that combining deep learning with probabilistic modeling, despite introducing a negligible temporal overhead, yields more efficient outcomes in terms of detection consistency.