Estimating the effect of age-induced defocus on object detection performance of automotive cameras

Defocus in automotive cameras can be caused by aging or changing temperature, and is one of the major reasons for degradation of data captured by the camera. This work investigates the effect of defocus on safety-critical object detection by comparing various pre-trained object detection models, primarily convolutional neural networks. These models are compared in their performance against static scenes with different levels of defocus. The static scene contained the three classes: person, bicycle, and cars, and the movement of the camera towards these objects was mimicked in discrete steps from 100 meter to 2 meter. During each distance step, the defocus was varied in 9 steps between +30 and -50 micrometer. To allow controlled defocusing, the general design of a state-of-the-art automotive camera was modified to allow lateral movement of the sensor in relation to the objective. All frames captured at 0 micrometer defocus were annotated for the three object classes, and these ground-truth annotations were compared to annotations from object detection models to calculate precision, recall and F1 score. The results showed that the precision remained close to 1.0 in multiple models, leaving the changes in F1 mainly impacted by changes in recall. Both large and small models performed well against defocus until 40 meter. At 60 meter the models performed within an acceptable range at low defocus and beyond 60 meter, recall dropped below 0.5. It was seen that at 100 meter the detection of objects was almost negligible, with the models detecting only one object or none.