FCPNet: A novel model to predict forward collision based-upon CNN

Predicting real-time collisions in unknown environments is a challenge especially for autonomous vehicles to avoid obstacles during missions. In this paper, we propose a novel approach called FCPNet which is a model based on Convolutional Neural Network (CNN) for a fast collision prediction. FCPNet is designed to run on CPU and embedded in an autonomous Unmanned Aerial Vehicle (UAV). Inspired by DroNet, FCPNet is composed of a ResNet-8 with three residual blocks, a dropout layer, and one output layer. Training and validation data are the Zurich images collected in an unstructured outdoor environment such as a city. This dataset consists of 31,972 color images in two classes: collision and no collision. FCPNet is compared to VGG-16, ResNet50V2, and MobileNetV2 models on CPU. The proposed model achieves the best F1-score and execution time on Raspberry Pi 4 thanks to the optimized parameters and class weights initialization to deal with an imbalanced dataset. The collision prediction probabilities as an alternative approach to ultrasonic sensors are presented in terms of confusion matrix showing the outstanding performance of FCPNet. We plan to use the proposed model with an obstacle avoidance policy to build robust and fast autonomous aerial vehicles.