Research on Low-Altitude Flight Obstacle Avoidance System for Police UAV Based on Multi-Sensor Fusion

With the increasing demand for urban security, police UAVs are facing the challenge of optimizing the collaborative performance of dynamic obstacle avoidance and covert penetration in complex low altitude environment. The current single-mode sensor system has significant limitations: under 105 DB electromagnetic interference, the positioning error of traditional millimeter wave radar exceeds 1.2 meters; In four concurrent dynamic threat scenarios, the collision probability of PID control strategy is 4.7%; The thermal radiation intensity is usually 30% higher than the environmental background. Although multi-sensor fusion technology improves the sensing accuracy through heterogeneous complementary features, in an area with a building density of 2800 square meters / square kilometers, spatio-temporal synchronization error and multipath effect lead to a 4.8-fold increase in the data conflict rate, thus limiting the effectiveness of covert penetration. The dynamic confidence fusion model proposed in this study breaks through the traditional theoretical framework of fixed weight fusion, and introduces a three-stage adaptive mechanism of real-time environmental interference assessment, sensor reliability feedback, and weight dynamic iteration for the first time. By establishing the quantitative mapping relationship between sensor confidence and electromagnetic interference intensity, motion state disturbance, the model realizes the independent enhancement of spatial-temporal correlation of heterogeneous data, and provides a new theoretical paradigm for solving multi-source data conflicts in a highly dynamic environment. The improved nash-q algorithm proposed in this study theoretically breaks through the convergence bottleneck of traditional reinforcement learning in multi-objective dynamic game: through the dynamic learning rate decay mechanism, it ensures that the strategy update converges to the ε - equilibrium point in a limited step; The Pareto optimal constraint term is introduced to compress the strategy space to a convergent convex subset to avoid the oscillation and divergence of traditional Q-learning in non cooperative games; The measured data show that the algorithm converges by an average of 15 steps in four threat concurrent scenarios, 58% faster than the standard nash-q. Experiments show that the cooperative design of rhombus scattering structure and thermoelectric active cooling system achieves an average radar cross section of -32.1 decibelsquare meters (DBSM) and suppresses the thermal radiation intensity to 12.8% of the environmental background. The dual-mode adaptive timing mechanism stabilizes the time alignment error at 1.5 nanoseconds (in which the quantum based timing is used for L5 strong interference scenes and the traditional timing is used for ordinary scenes), so that the trajectory planning accuracy exceeds the threshold of 0.15 meters. Under the combined conditions of L5 electromagnetic interference and wind speed of 22.4 M / s, the system achieved 83.2% computational performance and 78.5 kJ / km energy efficiency, which was 26.3% higher than the traditional solution. The research results provide a benchmark for low detectability flight control in complex scenes such as urban canyons, and promote the technological innovation of police UAVs in covert penetration missions.