Abstract:
Automatic Target Detection (ATD) and Recognition (ATR) from Thermal Infrared (TI)
imagery in defense and surveillance domain is a challenging computer vision (CV) task
in comparison to commercial autonomous vehicle perception domain. Peculiar domain
and TI modality challenges i.e., limited hardware, scale issues due to greater distances,
deliberate occlusion by tactical vehicles, limited sensor resolution and resultant lack of
structural information in targets, effects of weather, temperature and timeof- day variations and varying target to clutter ratios all result in increased intra-class variability and
reduced inter-class variability making real-time ATR a challenging CV task. Moreover,
owing to limited dataset availability for the TI modality, contemporary state-of-the-art
(SOTA) deep learning architecture for ATR largely remain under-explored. We propose a modified anchor-based single-stage detector called YOLOatr based on Yolov5s
with modifications to detection heads, feature-fusion in neck and a custom augmentation / hyperparameter profile. We evaluate the performance of our proposed model on
a comprehensive DSIAC MWIR dataset for real-time ATR over correlated and decorrelated testing protocols. Furthermore, we propose and validate a mode-driven deep
learning framework to reduce complexity in design, modification and implementation of
latest deep learning architectures with ease and simplicity with platform independent
modeling and automatic code-generation. We compare our results with leading studies
employing contemporary SOTA detectors. The result analysis demonstrates that our
proposed model achieves state-of-the-art performance compared to existing methods
with significant improvement in accuracy and detection speed
