Intelligent Perception and Situation Awareness for Automated vehicles

Overview & Synopsis

The Inria CHROMA research team develops new algorithms and methods for robotics, taking into consideration autonomy, limited resources, cooperation and social interactions. One of its main application domains are perception and decision making for driver assistance systems or autonomous cars. This work is done in the scope of long term collaborations with major automotive manufacturers such as Toyota and Renault. The approach for the perception of dynamic environment is based on sensor fusion and temporal filtering in occupancy grids using a probabilistic framework. The method is based on the merging of various sensor data into a probabilistic grid resulting in a joint estimation of the spatial occupancy and dynamics. It has been implemented using the Hybrid Sampling Bayesian Occupancy Filter paradigm (HSBOF) [1], and later extended with the Conditional Monte Carlo Dense Occupancy Tracker (CMCDOT) [2] to explicitly represent areas observability and efficiently extracts objects using a light-cost algorithm based on adaptive sampling of dynamic parts. This approach leads to greatly improve the quality of the results and to drastically decrease the computation and memory costs. The CMCDOT has been implemented and highly optimized using CUDA and embedded on several experimental platforms (Titan X, Jetson TK1, TX1) for performing real-time scene analysis. The outputs of this algorithm is used to perform risk estimation [3] and could be used in many intelligent vehicle applications such as emergency braking, obstacle avoidance or automatic driving. V2X communications between cars and infrastructures are also used to communicate objects position and to improve collision detection in blind spots or areas with low visibility. In this context, light connected perception units have been developed to be easily placed on dangerous road areas for warning connected cars of upcoming collisions. To perform experiments a Lexus LS 600h vehicle has been equipped with several sensors (2D Lidars, cameras, GPS, IMU) in cooperation with Toyota, as well as a Renault Zoe vehicle (3D and 2D Lidars, cameras, GPS, IMU) and the specifically designed Perception units (2D Lidars, cameras, Jetson TX1) in the scope of the French `Technological Research Institute Nanoelec'. References: [1] Amaury Negre, Lukas Rummelhard, and Christian Laugier. Hybrid Sampling Bayesian Occupancy Filter. In IEEE Intelligent Vehicles Symposium (IV), Dearborn, United States, Jun. 2014. [2] Lukas Rummelhard, Amaury Negre, and Christian Laugier. Conditional Monte Carlo Dense Occupancy Tracker. In 18th IEEE International Conference on Intelligent Transportation Systems, Las Palmas, Spain, Sep. 2015. [3] Lukas Rummelhard, Amaury Negre, Mathias Perrollaz, and Christian Laugier. Probabilistic Gridbased Collision Risk Prediction for Driving Application. In ISER, Marrakech/Essaouira, Morocco, Jun. 2014