Safe locomotion by AGV (Automated Guided Vehicle) in off-road terrains and unstructured environments requires robust execution of vision-mediated path navigation techniques. Efficient navigation can be obtained from precise surrounding perception and identification of manoeuvrable paths. However, the present well-known end-to-end approaches proved to have a substantial success rate of their performances in well-structured, on-road settings, but on the contrary, it is yet a challenge in complex and amorphous terrains to detect a proper manoeuvrable region for an autonomous point-to-point rescue or service robot. The research work majorly focuses on constructing an efficient perception system proposed as SegNav (Segmentation for Navigation). This is established featuring both semantic and geometric precision accuracy. The architecture of the projected model is constructed relative to reasonable insights of surroundings for obtaining consistent outputs. Feature extraction through group-wise attention defines the navigability of terrains. The assimilation of information is carried out by including multiview RGB inputs. The standardization of the model is established by training and experiment on a benchmarking off-road terrain dataset RUGD to evaluate its performance. It has also been tested on outdoor environmental custom data established from the real-world scenario. The model has also been compared with SOTA methods for benchmarking. SegNav achieves an improvement over mIoU (mean Intersection over Union) by 1.26%, and aAcc (Average Precision Accuracy) of 0.21% on the taken datasets. The findings of the experiment adds notable validation and confidence on AGV self-navigation across off road terrains.