Utilizing CNNs for Object Detection with LiDAR Data for Autonomous Driving

This project evaluates the feasibility of utilizing popular Convolutional Neural Networks (CNNs) to detect objects present in LiDAR (Light Detection And Ranging) data, and the resulting neural network's performance. This work aims to further existing experimentation using raw LiDAR data that is analyzed and represented in a two-dimensional frame. Using this method, hundreds of frames were generated to create a dataset that was used for neural network training and validation on an existing CNN architecture. The LiDAR dataset was used to train YOLOv3, a popular CNN model, to detect vehicles. This research aims to test a smaller version of the network, YOLOv3-tiny, to measure the change in accuracy between using YOLOv3 and YOLOv3-tiny on the LiDAR dataset. The results are then compared to the loss typically found when going from YOLOv3 to YOLOV3-tiny on camera-based images. In prior experimentation, a preprocessing method was also introduced to attempt to isolate target objects in the frame. The method will be evaluated in this paper to measure its effect on the final accuracy metric of the network. Lastly, the runtime performance of these networks will be evaluated on two embedded platforms to understand if the frame rate that the networks perform on is usable for real-world applications, based on the frame rate the sensor is capable of outputting and the inference speed of the network on the embedded platforms.