Publication

Abstract : To safely operate Level 4 autonomous vehicles, accurate and timely detection of cars and vans in urban areas is key. This paper presents a performance comparison of two 3DBEV object detection techniques based on point cloud data from the KITTI dataset: a hybrid model PIXOR vs. a complete model MV3D. Both models were trained using 5984 examples of vehicle point cloud data and tested with 1497. To provide for a fair evaluation, the model's point cloud was voxelized into a common format for use in constructing a BEV representation of the scene, and performance was evaluated using three metrics: recall, average precision, and mean average precision at standard intersection over union thresholds. The results show that MV3D has an average detection precision of 92% for cars and vans at short ranges while PIXOR has a slightly lower 91 % accuracy, but faster execution time. Because of its greater variability and stability of performance over time, PIXOR is more likely to be effective when performing real-time perception tasks than MV3D. While MV3D is slightly superior in some detection scenarios, it also requires a greater computational workload. These results demonstrate the tradeoff between accuracy and speed of inference for selecting suitable BEV object detection algorithms to use in developing reliable autonomous driving systems in highly populated urban settings.