Towards Robust Perception Depth Information For Collision Avoidance

Early detection of the obstacles and accurate estimation of the object’s distance helps avoid fatal accidents. However, the existing object detection ignores debris and other object classes not included in the training process. On the other hand, the driving area is monitored and recognized using active sensors like LiDAR, RADAR, but expensive. This study introduces a modified architecture to estimate the depth map based on an unsupervised learning framework. Furthermore, understanding the color-encoded depth map helps identify the risk of collision. The decoding of the color-encoded depth map provides information about the distance of the object. Thus, we presented an efficient and robust algorithm to predict a potential collision in real-time based on the estimated depth map using predefined threshold values. This approach emphasizes integrating the estimated depth map with the level of comprehension of the situation awareness to enhance the ability to recognize and process predicted uncertainties in the environment. The better results are achieved for modified architecture in terms of ARD, RMSE, RMSE (log), accuracy, and other evaluation metrics achieved lower but comparable results to the state-of-the-art techniques with a maximum depth of 80 meters. The integrated collision avoidance algorithm with the depth estimation architecture achieved a performance of 25 FPS on RTX 2080Ti GPU.