Structured Framework for Managing Decision Trade-offs in AI-Based Perception Systems for Advanced Driver Assistance Systems
The use of Advanced Driver Assistance Systems (ADAS) heavily depends on the perception models to make real-time decisions but the traditional methods have tended to use specific confidence thresholds to make the trade-offs between missed detections and false alarms to be not optimized. The proposed research establishes a risk-conscious perception and decision system that combines object detection with a mechanism of dynamically optimizing the threshold. YOLO based detection model is used to obtain bounding boxes and confidence scores of multi-view inputs (back, left and right cameras). The system considers a variety of confidence values and measures their effect by two safety-critical metrics: collision probability and false brake rate as opposed to using a fixed threshold. To give equal priority to these competing goals, a composite risk index is created where collision avoidance has a higher priority. The best threshold is one that will reduce this risk index in the decision space. The experimental analysis using a structured dataset reveals that the proposed framework has better safety-performance trade-offs than the traditional approaches. The model continues to minimize the risk of collision and low-level false alarms, which implies its applicability to a real-life setting when it comes to deploying an ADAS. The findings present the significance of adaptive decision strategies to safety-critical AI systems and offer an extensible basis to intelligent vehicle perception. Keywords: ADAS, Object Detection, YOLO, Risk-Aware Systems, Threshold Optimization, Collision Probability, False Alarm Rate, Intelligent Transportation Systems